diff --git a/DEVELOPMENT.md b/DEVELOPMENT.md
index 990f73bc..d57f243d 100644
--- a/DEVELOPMENT.md
+++ b/DEVELOPMENT.md
@@ -520,7 +520,8 @@ Copy the closest existing suite and modify. Required fields:
   "online_sla_ttft_ms": 500,
   "num_runs": 3,
   "warmup_runs": 1,
-  "online_warmup_runs": 0,
+  "online_warmup_requests": 10,
+  "burst_warmup_requests": 10,
   "interactive_warmup_runs": 0,
   "accuracy_threshold_delta": 0.1,
   "request_count": 200,
@@ -911,6 +912,71 @@ class InferenceResult:
 | speculative | Offline throughput with draft model (same path as offline, engine uses speculative decoding) | `throughput_tokens_per_sec`; optional `task.runtime_metrics.acceptance_rate` if runner overrides `get_runtime_metrics()` |
 | burst | Two-state bursty load: alternates steady QPS and burst QPS windows | `burst_degradation_ratio` (burst_ttft_p99 / steady_ttft_p99); `sla_met_during_burst` |
 
+### Warmup contract
+
+Cold engines inflate the first few timed requests by hundreds of ms (JIT
+compile, CUDA-graph allocation, KV cache priming). Each scenario discards
+a configurable prefix:
+
+| Scenario | Suite key | Default | Unit |
+|---|---|---|---|
+| offline / speculative / interactive | `warmup_runs` / `interactive_warmup_runs` | `1` / `0` | full passes |
+| online | `online_warmup_requests` | `10` | dummy requests fired before QPS sweep |
+| burst | `burst_warmup_requests` | `10` | dummy requests fired before first cycle |
+| sustained | `warmup_minutes` | `2` | minutes of samples excluded from analysis |
+
+Warmup-time exceptions are logged and swallowed — they never abort the
+timed phase.
+
+### Reliability metrics
+
+Each scenario emits an inter-run reliability block alongside its primary
+metrics so submitters can prove their results are reproducible without
+shipping `samples.jsonl`. Shape:
+
+```json
+{
+  "n":         3,
+  "mean":      1234.5,
+  "std":         21.3,
+  "cv_pct":      1.7,
+  "stability": "stable",
+  "runs": [1230.1, 1255.2, 1218.2]
+}
+```
+
+`stability` thresholds: `cv_pct ≤ 3 → stable ✓`, `≤ 8 → noisy ⚠`,
+otherwise `high-variance`. Calibrated from the May-2026 backfill — see
+the comment above `_STABILITY_THRESHOLD_*` in `loadgen/loadgen.py` for the
+empirical distribution that informed the choice. Tunable centrally there.
+
+**`high-variance` is informational, not a verdict.** High CV means the
+hardware × workload combo carries irreducible jitter (thermal throttle on
+consumer cards, HCCL noise on 16-chip Ascend topologies, acceptance-rate
+fluctuation on speculative decoding) — it is **not** a sign the
+submission is broken. The frontend reflects this: high-variance pills
+use an orange tone with no error glyph, while only stable / noisy carry
+✓ / ⚠ icons.
+
+If you submit a result that lands as high-variance, you do not need to
+re-run. The badge is for downstream readers picking hardware for
+latency-sensitive workloads — they can use the CV % to size their
+safety margins, while peak-throughput shoppers can largely ignore it.
+
+| Scenario | Field path | Reliability source |
+|---|---|---|
+| offline | `metrics.offline.results_by_concurrency[i].throughput_tokens_per_sec_reliability` | per-run throughput across `num_runs` |
+| online | `metrics.online.results_by_qps[i].ttft_ms_p99_reliability` | per-run TTFT p99 across `num_runs` |
+| interactive | `metrics.interactive.ttft_ms_p99_reliability` | per-run TTFT p99 across `num_runs` |
+| sustained | `metrics.sustained.throughput_post_warmup_reliability` | per-interval throughput (post-warmup) |
+| burst | `metrics.burst.recovery_time_seconds` (+ `_per_cycle`) | seconds until rolling p99 returns to ≤ 1.5× steady baseline |
+
+Backfilling these for existing results is done by
+`tools/backfill_distribution_stats.py`, which reads each result's local
+`samples.jsonl` and writes the summary stats in place. Offline reliability
+cannot be backfilled because per-run throughput was never recorded in
+`samples.jsonl` historically — it stays `{}` for old offline results.
+
 ---
 
 ## Schema and Validation
diff --git a/leaderboard/generate.py b/leaderboard/generate.py
index bf35de21..ead41d62 100644
--- a/leaderboard/generate.py
+++ b/leaderboard/generate.py
@@ -6,6 +6,8 @@
     python leaderboard/generate.py
 """
 
+from __future__ import annotations
+
 import hashlib
 import json
 import re
@@ -238,6 +240,11 @@ def extract_detail(result: dict) -> dict:
         "meta_model_load_sec":   meta.get("model_load_seconds"),
         "meta_start_time":       meta.get("benchmark_start_time"),
         "meta_notes":            meta.get("notes"),
+        # Vendor-specific environment fields collected by platforms/<vendor>.py
+        # (e.g. ROCm-SMI link health, NVML clock telemetry). The modal flattens
+        # this dict and shows only non-null entries — different vendors record
+        # different keys by design and no UI tries to unify them.
+        "env_vendor_details":    env.get("vendor_details") or {},
     }
 
 
@@ -297,6 +304,9 @@ def _concurrency_labels(rows):
     def _online_block():
         online   = metrics.get("online", {})
         qps_rows = online.get("results_by_qps", [])
+        # Per-QPS reliability blocks. Emitted as a parallel array so the
+        # frontend can render a badge next to each QPS row without joining
+        # by index from a separate object.
         return {
             "labels":        [str(r.get("target_qps", "")) for r in qps_rows],
             "ttft_p50":      [r.get("ttft_ms_p50") for r in qps_rows],
@@ -304,6 +314,8 @@ def _online_block():
             "tpot_p50":      [r.get("tpot_ms_p50") for r in qps_rows],
             "sla_met":       [r.get("sla_met")      for r in qps_rows],
             "max_valid_qps": online.get("max_valid_qps"),
+            "ttft_ms_p99_reliability":
+                [r.get("ttft_ms_p99_reliability") or {} for r in qps_rows],
         }
 
     def _interactive_block():
@@ -315,6 +327,7 @@ def _interactive_block():
             "tpot_p50": iv.get("tpot_ms_p50"),
             "tpot_p90": iv.get("tpot_ms_p90"),
             "tpot_p99": iv.get("tpot_ms_p99"),
+            "ttft_ms_p99_reliability": iv.get("ttft_ms_p99_reliability") or {},
         }
 
     def _sustained_block():
@@ -334,6 +347,8 @@ def _sustained_block():
             "throttle_ratio":        s.get("throttle_ratio"),
             "throttle_onset_minute": s.get("throttle_onset_minute"),
             "ttft_p99_drift_ms":     s.get("ttft_p99_drift_ms"),
+            "throughput_post_warmup_reliability":
+                s.get("throughput_post_warmup_reliability") or {},
             "samples":               samples,
         }
 
@@ -352,6 +367,9 @@ def _burst_block():
             "burst_requests_total":         b.get("burst_requests_total"),
             "sla_met_during_burst":         b.get("sla_met_during_burst"),
             "burst_degradation_ratio":      b.get("burst_degradation_ratio"),
+            "recovery_time_seconds":        b.get("recovery_time_seconds"),
+            "recovery_time_seconds_per_cycle":
+                b.get("recovery_time_seconds_per_cycle") or [],
             "results_by_cycle":             b.get("results_by_cycle"),
         }
 
@@ -370,6 +388,11 @@ def _speculative_block():
             "mean_accepted_tokens": rm.get("mean_accepted_tokens"),
         }
 
+    # Per-concurrency-level offline reliability blocks. Parallel array to
+    # `throughput` and `memory_gb` so the frontend can join by row index.
+    def _offline_reliability(rows):
+        return [r.get("throughput_tokens_per_sec_reliability") or {} for r in rows]
+
     if suite == "suite_A":
         rows = _offline_rows()
         return {
@@ -378,6 +401,7 @@ def _speculative_block():
                 "labels":     _concurrency_labels(rows),
                 "throughput": [r.get("throughput_tokens_per_sec") for r in rows],
                 "memory_gb":  [r.get("peak_memory_gb")            for r in rows],
+                "throughput_reliability": _offline_reliability(rows),
             },
             "online":      _online_block(),
             "interactive": _interactive_block(),
@@ -395,6 +419,7 @@ def _speculative_block():
                 "throughput":          [r.get("throughput_tokens_per_sec")          for r in rows],
                 "throughput_per_chip": [r.get("throughput_tokens_per_sec_per_chip") for r in rows],
                 "memory_gb":           [r.get("peak_memory_gb")                     for r in rows],
+                "throughput_reliability": _offline_reliability(rows),
             },
             "online":    _online_block(),
             "sustained": _sustained_block(),
@@ -409,6 +434,7 @@ def _speculative_block():
                 "labels":     _concurrency_labels(rows),
                 "throughput": [r.get("throughput_tokens_per_sec") for r in rows],
                 "memory_gb":  [r.get("peak_memory_gb")            for r in rows],
+                "throughput_reliability": _offline_reliability(rows),
             },
             "interactive": _interactive_block(),
             "sustained":   _sustained_block(),
@@ -514,6 +540,7 @@ def _speculative_block():
                 "labels":     _concurrency_labels(rows),
                 "throughput": [r.get("throughput_tokens_per_sec") for r in rows],
                 "memory_gb":  [r.get("peak_memory_gb")            for r in rows],
+                "throughput_reliability": _offline_reliability(rows),
             },
             "online":      _online_block(),
             "interactive": _interactive_block(),
@@ -530,6 +557,7 @@ def _speculative_block():
                 "labels":     _concurrency_labels(rows),
                 "throughput": [r.get("throughput_tokens_per_sec") for r in rows],
                 "memory_gb":  [r.get("peak_memory_gb")            for r in rows],
+                "throughput_reliability": _offline_reliability(rows),
             },
             "online":      _online_block(),
             "interactive": _interactive_block(),
diff --git a/leaderboard/site/assets/css/modal.css b/leaderboard/site/assets/css/modal.css
index 00e8e21a..7b456224 100644
--- a/leaderboard/site/assets/css/modal.css
+++ b/leaderboard/site/assets/css/modal.css
@@ -127,6 +127,79 @@
   margin-right: 0.3rem;
 }
 
+/* Inter-run reliability pill that lives in the modal subtitle. Colours
+ * track --good / --warn / --bad so the existing palette controls dark-mode
+ * behaviour. We intentionally tone down opacity so the badge does not
+ * compete with the primary metric callouts above. */
+.modal-reliab-pill {
+  display: inline-flex;
+  align-items: center;
+  gap: 0.25rem;
+  padding: 0.1rem 0.5rem;
+  border-radius: 999px;
+  font-size: 0.7rem;
+  font-weight: 600;
+  letter-spacing: 0.01em;
+  border: 1px solid color-mix(in srgb, currentColor 35%, transparent);
+  background: color-mix(in srgb, currentColor 10%, transparent);
+  /* Pill is rendered as a <button> so it can be clicked / focused; strip
+   * the default button chrome and signal "this is hoverable for help"
+   * via the cursor + the ⓘ glyph inside. */
+  font-family: inherit;
+  cursor: help;
+  transition: filter 0.15s ease, transform 0.15s ease;
+}
+.modal-reliab-pill:hover  { filter: brightness(1.05); transform: translateY(-1px); }
+.modal-reliab-pill:active { transform: translateY(0); }
+.modal-reliab-pill:focus-visible {
+  outline: 2px solid color-mix(in srgb, currentColor 55%, transparent);
+  outline-offset: 2px;
+}
+.modal-reliab-help {
+  opacity: 0.65;
+  font-weight: 500;
+  font-size: 0.78em;
+}
+.modal-reliab-pill.stable         { color: var(--good, #2da44e); }
+.modal-reliab-pill.noisy          { color: var(--warn, #d29922); }
+/* High-variance uses an orange tone (not pure red) on purpose: red would
+ * read as an error glyph, but a high-CV submission may be a perfectly
+ * correct measurement of a legitimately jittery hardware × workload pair
+ * (e.g. consumer cards under sustained load). Orange says "look at this"
+ * without saying "this is broken". */
+.modal-reliab-pill.high_variance  { color: #d97706; }
+.modal-reliab-pill.unknown        { color: var(--fg-faint); }
+
+/* Inline help paragraph rendered under a detail-section title. Currently
+ * only the Reliability section uses this, so the styling stays modest —
+ * we don't want it competing with the data rows below. */
+.detail-section-help {
+  margin: 0 0 0.55rem 0;
+  font-size: 0.78rem;
+  line-height: 1.5;
+  color: var(--fg-muted, #6b7280);
+}
+.detail-section-help .reliab-legend {
+  display: inline-block;
+  padding: 0 0.3rem;
+  border-radius: 4px;
+  font-weight: 600;
+  font-size: 0.92em;
+  background: color-mix(in srgb, currentColor 12%, transparent);
+}
+.detail-section-help .reliab-legend.stable        { color: var(--good, #2da44e); }
+.detail-section-help .reliab-legend.noisy         { color: var(--warn, #d29922); }
+.detail-section-help .reliab-legend.high_variance { color: #d97706; }
+
+/* Brief highlight flash when a click-jump from the subtitle pill lands on
+ * a detail section, so the user notices the scroll target. The animation
+ * is deliberately gentle (1.4 s fade-out via JS-toggled class). */
+.detail-section.detail-section-flash {
+  background: color-mix(in srgb, var(--accent, #2563eb) 8%, transparent);
+  border-radius: 6px;
+  transition: background-color 0.4s ease;
+}
+
 .modal-close {
   background: transparent;
   border: 1px solid var(--border-soft);
diff --git a/leaderboard/site/assets/js/modal.js b/leaderboard/site/assets/js/modal.js
index 3461dbbf..151d759a 100644
--- a/leaderboard/site/assets/js/modal.js
+++ b/leaderboard/site/assets/js/modal.js
@@ -114,6 +114,16 @@ export function initModal() {
       _downloadVizChart(dlBtn);
       return;
     }
+    // Pills / buttons that ask the modal to scroll to a named section
+    // inside the Details panel. Currently used by the reliability pill in
+    // the subtitle; future header chips can reuse `data-scroll-to` for
+    // the same effect without each one wiring its own handler.
+    const jumpTrigger = ev.target.closest("[data-scroll-to]");
+    if (jumpTrigger && _modalEl.contains(jumpTrigger)) {
+      ev.preventDefault();
+      _scrollToDetailSection(jumpTrigger.dataset.scrollTo);
+      return;
+    }
   });
 
   // Esc closes from anywhere.
@@ -293,6 +303,8 @@ function _fillModal(row) {
     parts.push(esc(row.precision) + fallback);
   }
   if (row.date) parts.push(esc(fmtDate(row.date)));
+  const pill = _reliabilityPill(row);
+  if (pill) parts.push(pill);
   subEl.innerHTML = parts.filter(Boolean).join(' <span class="sep">·</span> ');
 
   // Footer: submission + script link.
@@ -381,17 +393,250 @@ function _detailRow(label, v, opts = {}) {
   return `<div class="detail-row"><span class="detail-label">${esc(label)}</span><span class="detail-value${mono}${cls}">${display}</span></div>`;
 }
 
-function _detailSection(title, rows) {
+function _detailSection(title, rows, opts = {}) {
   const content = rows.filter(Boolean).join("");
   if (!content) return "";
+  // `anchor` becomes a data attribute the subtitle pill can scroll to.
+  // `caption` is a small prose block rendered under the section title —
+  // used by Reliability to explain CV / thresholds inline so readers
+  // don't need to rely on the (delayed, plain-text) native title tooltip.
+  const anchor = opts.anchor ? ` data-section="${esc(opts.anchor)}"` : "";
+  const caption = opts.caption
+    ? `<p class="detail-section-help">${opts.caption}</p>`
+    : "";
   return `
-    <div class="detail-section">
+    <div class="detail-section"${anchor}>
       <div class="detail-section-title">${esc(title)}</div>
+      ${caption}
       ${content}
     </div>
   `;
 }
 
+// Scroll the Details tab body to the section tagged data-section=<name>.
+// Activates the Details tab first if the user is on Viz / Implementation
+// when they click the subtitle pill, so the scroll target is actually
+// in the visible panel.
+function _scrollToDetailSection(name) {
+  if (!_modalEl || !name) return;
+  if (_activeTab !== "details") _setTab("details");
+  const target = _modalEl.querySelector(`.detail-section[data-section="${name}"]`);
+  if (!target) return;
+  // Use the modal body as the scroll container so we don't reflow the
+  // whole document. `scrollIntoView` with `{ block: "start" }` would push
+  // the section to the very top, hiding the title — offset by 12 px so
+  // the section heading stays nicely below the modal subheader.
+  const body = _modalEl.querySelector(".modal-body");
+  if (body) {
+    const top = target.offsetTop - 12;
+    body.scrollTo({ top, behavior: "smooth" });
+  } else {
+    target.scrollIntoView({ behavior: "smooth", block: "start" });
+  }
+  // Brief highlight so users notice where the page jumped to.
+  target.classList.add("detail-section-flash");
+  setTimeout(() => target.classList.remove("detail-section-flash"), 1400);
+}
+
+// ── Reliability rendering ────────────────────────────────────────────────────
+//
+// Reliability blocks are emitted by loadgen.py for each scenario starting in
+// the "feat: emit reliability stats" series. Each block has shape:
+//   { n, mean, std, cv_pct, stability: "stable"|"noisy"|"high-variance", runs:[…] }
+// Older results (pre-feature) carry an empty {} which we skip silently — the
+// section header is suppressed if no scenario reported a block, so old runs
+// look the same as before.
+//
+// Design note on iconography:
+//   stable          → ✓  Green — the headline number is the property of the chip.
+//   noisy           → ⚠  Amber — modest jitter (≤8 % CV). Common, not a problem.
+//   high-variance   →    No glyph, red-tinted text — informational, not a verdict.
+// We intentionally do NOT use an ✗ for high-variance. High CV means "this
+// hardware × workload combo carries irreducible variability the reader
+// should be aware of", not "this measurement is wrong". An ✗ would
+// implicitly indict submissions whose hardware simply lacks proper cooling,
+// or scale-out runs whose network jitter is genuine. The colour cue
+// communicates "look closer" without the verdict.
+
+const _STABILITY_ICON = { stable: "✓", noisy: "⚠", "high-variance": "" };
+
+function _hasReliability(block) {
+  return block && typeof block === "object" && block.cv_pct != null;
+}
+
+// Pick the worst (highest CV%) reliability block from a row's viz, used by
+// the modal subtitle pill. Returns { cv_pct, stability, scenario, metric } or
+// null. The "scenario" hint helps users find where to drill in.
+function _pickWorstReliability(row) {
+  const viz = row.viz || {};
+  let worst = null;
+  const consider = (b, scenario, metric) => {
+    if (!_hasReliability(b)) return;
+    if (!worst || b.cv_pct > worst.cv_pct) {
+      worst = { cv_pct: b.cv_pct, stability: b.stability, scenario, metric };
+    }
+  };
+
+  // offline (per-concurrency, multiple blocks)
+  if (viz.offline && Array.isArray(viz.offline.throughput_reliability)) {
+    viz.offline.throughput_reliability.forEach((b) =>
+      consider(b, "offline", "throughput"));
+  }
+  // online (per-QPS, multiple blocks)
+  if (viz.online && Array.isArray(viz.online.ttft_ms_p99_reliability)) {
+    viz.online.ttft_ms_p99_reliability.forEach((b) =>
+      consider(b, "online", "TTFT p99"));
+  }
+  // interactive (single block)
+  consider(viz.interactive && viz.interactive.ttft_ms_p99_reliability,
+           "interactive", "TTFT p99");
+  // sustained (single block)
+  consider(viz.sustained && viz.sustained.throughput_post_warmup_reliability,
+           "sustained", "throughput");
+  return worst;
+}
+
+// Build the small `cv X.X% ✓` pill that appears in the modal subtitle row.
+//
+// The pill is interactive in three ways, layered so each user type has a
+// path to the full explanation:
+//   - Hover (mouse)        → native title tooltip with one-line explanation
+//   - Click (any device)   → scrolls the Details tab to the Reliability
+//                            section, which carries the full prose caption
+//   - `cursor: help` + ⓘ glyph cue readers that this is hoverable
+//
+// The data-scroll-to attribute lets us delegate the click handler at the
+// modal level (see _wireReliabilityPillJump) — no per-pill listener needed.
+function _reliabilityPill(row) {
+  const w = _pickWorstReliability(row);
+  if (!w) return "";
+  const icon = _STABILITY_ICON[w.stability] || "";
+  // CSS class names must be CSS-safe: replace "-" → "_" for high_variance.
+  const stabilitySlug = (w.stability || "unknown").replace(/-/g, "_");
+  const cls  = "modal-reliab-pill " + stabilitySlug;
+  const title = `Coefficient of variation across runs: ${w.cv_pct}% ` +
+    `(worst across scenarios; from ${w.scenario} ${w.metric}). ` +
+    `≤3% stable, ≤8% noisy, >8% high-variance. ` +
+    `High-variance is informational — it means this hardware × workload ` +
+    `combo has irreducible jitter, not that the measurement is wrong. ` +
+    `Click for details.`;
+  const head = icon ? `${icon} cv ${w.cv_pct}%` : `cv ${w.cv_pct}%`;
+  return `<button type="button" class="${esc(cls)}" ` +
+         `title="${esc(title)}" data-scroll-to="reliability">` +
+         `reliability ${esc(head)} <span class="modal-reliab-help" aria-hidden="true">ⓘ</span>` +
+         `</button>`;
+}
+
+// Render one row per scenario in the Details tab. Skipped if a scenario has
+// no block (older results). Burst gets recovery_time_seconds appended.
+function _reliabilityRows(row) {
+  const viz = row.viz || {};
+  const rows = [];
+
+  // Format: "4.19% ⚠ noisy (n=14)" or "12.4% high-variance (n=14)".
+  // Icon is omitted (not just blank) for high-variance so we don't render
+  // a dangling "·" — the label and colour carry the meaning instead.
+  const fmtBlock = (b) => {
+    const icon = _STABILITY_ICON[b.stability] || "";
+    const head = icon ? `${b.cv_pct}% · ${icon}` : `${b.cv_pct}%`;
+    return `${head} <span class="muted-note">${esc(b.stability || "")} (n=${b.n})</span>`;
+  };
+
+  // offline — show the worst (largest CV) of all client_concurrency rows.
+  // That's the limiting concurrency for stability claims.
+  if (viz.offline && Array.isArray(viz.offline.throughput_reliability)) {
+    const labels = viz.offline.labels || [];
+    const blocks = viz.offline.throughput_reliability;
+    const indexed = blocks
+      .map((b, i) => ({ b, label: labels[i] || `cc=${i}` }))
+      .filter((x) => _hasReliability(x.b));
+    if (indexed.length) {
+      indexed.sort((a, b) => b.b.cv_pct - a.b.cv_pct);
+      const w = indexed[0];
+      rows.push(_detailRow(
+        `Offline throughput (cc=${w.label})`,
+        fmtBlock(w.b),
+        { html: true },
+      ));
+    }
+  }
+
+  if (viz.online && Array.isArray(viz.online.ttft_ms_p99_reliability)) {
+    const labels = viz.online.labels || [];
+    const blocks = viz.online.ttft_ms_p99_reliability;
+    const indexed = blocks
+      .map((b, i) => ({ b, label: labels[i] || `qps=${i}` }))
+      .filter((x) => _hasReliability(x.b));
+    if (indexed.length) {
+      indexed.sort((a, b) => b.b.cv_pct - a.b.cv_pct);
+      const w = indexed[0];
+      rows.push(_detailRow(
+        `Online TTFT p99 (qps=${w.label})`,
+        fmtBlock(w.b),
+        { html: true },
+      ));
+    }
+  }
+
+  if (viz.interactive && _hasReliability(viz.interactive.ttft_ms_p99_reliability)) {
+    rows.push(_detailRow(
+      "Interactive TTFT p99",
+      fmtBlock(viz.interactive.ttft_ms_p99_reliability),
+      { html: true },
+    ));
+  }
+
+  if (viz.sustained && _hasReliability(viz.sustained.throughput_post_warmup_reliability)) {
+    rows.push(_detailRow(
+      "Sustained throughput (post-warmup)",
+      fmtBlock(viz.sustained.throughput_post_warmup_reliability),
+      { html: true },
+    ));
+  }
+
+  // Burst — non-CV stability metric: time-to-recover after a peak window.
+  if (viz.burst && viz.burst.recovery_time_seconds != null) {
+    const sec = Number(viz.burst.recovery_time_seconds);
+    rows.push(_detailRow(
+      "Burst recovery time",
+      `${sec.toFixed(2)} s ` +
+      `<span class="muted-note">median per-cycle, threshold 1.5× steady p99</span>`,
+      { html: true },
+    ));
+  } else if (viz.burst && Array.isArray(viz.burst.recovery_time_seconds_per_cycle)
+             && viz.burst.recovery_time_seconds_per_cycle.length === 0) {
+    // Burst ran but never recovered within any cycle's post-burst window.
+    rows.push(_detailRow(
+      "Burst recovery time",
+      `not measurable <span class="muted-note">(never returned to baseline within steady window)</span>`,
+      { html: true },
+    ));
+  }
+
+  return rows;
+}
+
+// Flatten env_info.vendor_details into rows. Keys that are objects/arrays are
+// JSON-stringified for display; null/empty values are dropped. We intentionally
+// do not try to humanise key names — vendors disagree on terminology and the
+// keys themselves are the documented contract.
+function _vendorDetailRows(row) {
+  const obj = (row.detail || {}).env_vendor_details;
+  if (!obj || typeof obj !== "object") return [];
+  const rows = [];
+  for (const k of Object.keys(obj).sort()) {
+    const v = obj[k];
+    if (v === null || v === undefined || v === "") continue;
+    if (Array.isArray(v) && v.length === 0) continue;
+    if (typeof v === "object" && !Array.isArray(v) && Object.keys(v).length === 0) continue;
+    const display = (typeof v === "object")
+      ? JSON.stringify(v)
+      : String(v);
+    rows.push(_detailRow(k, display, { mono: true }));
+  }
+  return rows;
+}
+
 function _renderDetails(row, panel) {
   const d = row.detail || {};
 
@@ -457,6 +702,21 @@ function _renderDetails(row, panel) {
       d.run_pp != null ? _detailRow("Pipeline parallel size", d.run_pp) : null,
       d.run_dp != null ? _detailRow("Data parallel size",     d.run_dp) : null,
     ]),
+    _detailSection("Reliability", _reliabilityRows(row), {
+      anchor: "reliability",
+      // Plain HTML (not auto-escaped) so we can highlight the threshold
+      // pills. Keep the wording short — readers shouldn't need to read a
+      // paragraph to decode a single percentage in the table below.
+      caption:
+        "Inter-run <strong>coefficient of variation</strong> " +
+        "(CV = std / mean × 100%) across the runs that produced this " +
+        "submission — lower is more reproducible. " +
+        "<span class=\"reliab-legend stable\">✓ stable</span> ≤ 3% &nbsp;·&nbsp; " +
+        "<span class=\"reliab-legend noisy\">⚠ noisy</span> ≤ 8% &nbsp;·&nbsp; " +
+        "<span class=\"reliab-legend high_variance\">high-variance</span> &gt; 8% " +
+        "(informational — natural jitter, not a measurement error).",
+    }),
+    _detailSection("Vendor-specific environment", _vendorDetailRows(row)),
     _detailSection("Accuracy", [
       _detailRow("Subset score", d.acc_score, {
         format: (v) => Number(v).toFixed(2),
diff --git a/loadgen/loadgen.py b/loadgen/loadgen.py
index 758abc41..7a917108 100644
--- a/loadgen/loadgen.py
+++ b/loadgen/loadgen.py
@@ -66,6 +66,123 @@ def _percentile(data: list, p: float):
     return sorted_data[lo] + (sorted_data[hi] - sorted_data[lo]) * (idx - lo)
 
 
+# ── Reliability helpers ──────────────────────────────────────────────────────
+#
+# These produce the inter-run variability metrics consumed by the leaderboard
+# UI's "Reliability" panel. They live here (not in types.py) because they are
+# pure functions over already-collected per-run lists and are easier to
+# regression-test alongside the scenario implementations.
+#
+# Coefficient of Variation (CV) = std / mean × 100 %, computed with ddof=1
+# (sample std) when n ≥ 2. Returns None when input is too small or the mean
+# is non-positive, in which case the frontend hides the badge entirely so
+# users do not see a meaningless "stable ✓" on a single-run measurement.
+
+# Stability thresholds. Calibrated from the initial 255-result sustained
+# backfill (May 2026), which had a CV median of 3.1 % and p90 of 13.1 %.
+# Tighter thresholds (e.g. ≤ 2 % / ≤ 5 %) labelled the literal median run
+# "noisy" and ~30 % of submissions "unstable" — those labels were too
+# pejorative for what is really normal hardware jitter. The buckets here
+# split the empirical distribution into ~48 % stable / ~35 % noisy /
+# ~17 % high-variance, with the high-variance bucket dominated by chips
+# we expect to genuinely throttle (RTX 5090, A6000, V100s) or scale-out
+# topologies with real network jitter (Ascend ×16).
+#
+# Important wording choice: the third tier is named "high-variance", not
+# "unstable", because high CV does not mean the measurement is wrong — it
+# means the headline number carries irreducible variability the reader
+# should be aware of. The frontend reflects this with a colour cue and
+# no error glyph; "high-variance" is a description, not a verdict.
+_STABILITY_THRESHOLD_STABLE_PCT = 3.0
+_STABILITY_THRESHOLD_NOISY_PCT  = 8.0
+
+
+def _cv_pct(values: list) -> Optional[float]:
+    """Coefficient of variation as a percentage. None if too small / undefined."""
+    if not values or len(values) < 2:
+        return None
+    arr = np.asarray(values, dtype=float)
+    arr = arr[np.isfinite(arr)]
+    if len(arr) < 2:
+        return None
+    mean = float(arr.mean())
+    if mean <= 0:
+        return None
+    std = float(arr.std(ddof=1))
+    return round(std / mean * 100.0, 2)
+
+
+def _stability_label(cv_pct: Optional[float]) -> Optional[str]:
+    """Map a CV percentage to a stable / noisy / high-variance label, or None."""
+    if cv_pct is None:
+        return None
+    if cv_pct <= _STABILITY_THRESHOLD_STABLE_PCT:
+        return "stable"
+    if cv_pct <= _STABILITY_THRESHOLD_NOISY_PCT:
+        return "noisy"
+    return "high-variance"
+
+
+def _reliability_block(values: list, *, decimals: int = 2) -> dict:
+    """
+    Build the standard {n, mean, std, cv_pct, stability, runs} block emitted
+    per metric. Returns an empty dict (not None) so the result schema retains
+    a consistent shape — frontend gates on `cv_pct` being numeric.
+    """
+    if not values:
+        return {}
+    arr = np.asarray(values, dtype=float)
+    arr = arr[np.isfinite(arr)]
+    if len(arr) == 0:
+        return {}
+    mean = float(arr.mean())
+    std  = float(arr.std(ddof=1)) if len(arr) >= 2 else 0.0
+    cv   = _cv_pct(arr.tolist())
+    return {
+        "n":         int(len(arr)),
+        "mean":      round(mean, decimals),
+        "std":       round(std, decimals),
+        "cv_pct":    cv,
+        "stability": _stability_label(cv),
+        "runs":      [round(float(v), decimals) for v in arr.tolist()],
+    }
+
+
+def _compute_recovery_time(
+    arrivals: list,
+    ttfts: list,
+    *,
+    threshold_ms: float,
+    window_s: float = 3.0,
+    min_samples: int = 5,
+) -> Optional[float]:
+    """
+    Find the elapsed time (seconds, relative to the start of the post-burst
+    steady window) at which a rolling-window p99 of TTFT first falls below
+    `threshold_ms`. Returns None if it never recovers within the window or
+    if there are too few samples to compute a stable percentile.
+
+    `arrivals` and `ttfts` are parallel arrays — arrivals must be relative
+    times in seconds from the start of the measurement window.
+    """
+    if not arrivals or len(arrivals) < min_samples:
+        return None
+    pairs = sorted(zip(arrivals, ttfts))
+    a = [p[0] for p in pairs]
+    t = [p[1] for p in pairs]
+    n = len(a)
+    j = 0
+    for i in range(n):
+        while j < i and a[j] < a[i] - window_s:
+            j += 1
+        if i - j + 1 < min_samples:
+            continue
+        window = t[j:i + 1]
+        if float(np.percentile(window, 99)) < threshold_ms:
+            return round(float(a[i]), 2)
+    return None
+
+
 class AccelMarkLoadGen:
 
     def __init__(
@@ -95,19 +212,35 @@ def __init__(
         # Use different request counts per scenario
         # offline: use request_count (default 200, fast)
         # online/interactive: use online_request_count if set, else all requests
+        #
+        # Warmup semantics differ per scenario:
+        #   offline / interactive : `warmup_runs` = number of full passes to discard
+        #                           (interactive_warmup_runs may override for interactive)
+        #   sustained             : `warmup_minutes` = time window discarded
+        #   online / burst        : `online_warmup_requests` / `burst_warmup_requests`
+        #                           = number of dummy requests fired sequentially before
+        #                           the timed phase, used to JIT-compile kernels, allocate
+        #                           CUDA graphs, prime the KV cache, etc. Results are
+        #                           never recorded. Without this warmup, the first few
+        #                           requests of the first QPS level inflate p99 by
+        #                           hundreds of ms on cold engines.
+        self.online_warmup_requests = 0
+        self.burst_warmup_requests = 0
         if scenario == "offline":
             count = suite.get("request_count")
             self.warmup_runs = suite.get("warmup_runs", 1)
         elif scenario == "online":
             # online and interactive need more requests for reliable p99
             count = suite.get("online_request_count", suite.get("request_count"))
-            self.warmup_runs = suite.get("online_warmup_runs", 0)
+            self.warmup_runs = 0  # online doesn't use full-pass warmup
+            self.online_warmup_requests = suite.get("online_warmup_requests", 10)
         elif scenario == "interactive":
             count = suite.get("interactive_request_count", suite.get("request_count"))
             self.warmup_runs = suite.get("interactive_warmup_runs", 0)
         elif scenario == "burst":
             count = suite.get("online_request_count", suite.get("request_count"))
-            self.warmup_runs = suite.get("online_warmup_runs", 0)
+            self.warmup_runs = 0  # burst doesn't use full-pass warmup
+            self.burst_warmup_requests = suite.get("burst_warmup_requests", 10)
         elif scenario == "speculative":
             count = suite.get("request_count")
             self.warmup_runs = suite.get("warmup_runs", 1)
@@ -138,8 +271,7 @@ def run(self, inference_fn: Callable) -> dict:
                     "_run_interactive requires an async inference_fn(request: InferenceRequest) -> InferenceResult. "
                     "Pass an async coroutine (inference_fn_streaming)."
                 )
-            loop = asyncio.get_event_loop()
-            return loop.run_until_complete(self._run_interactive_async(inference_fn))
+            return asyncio.run(self._run_interactive_async(inference_fn))
         elif self.scenario == "training":
             return self._run_training(inference_fn)
         elif self.scenario == "multiturn":
@@ -324,6 +456,12 @@ def _fire_request() -> None:
         if len(ttft_p99s) >= 2:
             ttft_p99_drift_ms = round(ttft_p99s[-1] - ttft_p99s[0], 1)
 
+        # Inter-sample throughput stability across the post-warmup window.
+        # This is conceptually distinct from `throttle_ratio` (min/max): CV
+        # measures dispersion around the mean and is a better signal for
+        # "the chip throttles intermittently" vs "the chip is degrading".
+        throughput_cv_block = _reliability_block(throughputs, decimals=1)
+
         return {
             "sustained": {
                 "sustained_concurrency":               sustained_concurrency,
@@ -335,6 +473,7 @@ def _fire_request() -> None:
                 "throttle_ratio":                      throttle_ratio,
                 "throttle_onset_minute":               throttle_onset_minute,
                 "ttft_p99_drift_ms":                   ttft_p99_drift_ms,
+                "throughput_post_warmup_reliability":  throughput_cv_block,
             }
         }
 
@@ -502,6 +641,12 @@ def _run_offline(self, inference_fn: Callable) -> dict:
                     "power_watts_avg": None,
                     "power_watts_peak": None,
                     "oom": False,
+                    # Per-run throughput reliability: lets the UI show "stable ✓ /
+                    # noisy ⚠ / unstable ✗" without forcing the user to download
+                    # samples.jsonl. `runs` preserves the underlying values so
+                    # future stability rules can be recomputed without a re-run.
+                    "throughput_tokens_per_sec_reliability":
+                        _reliability_block(run_throughputs, decimals=2),
                     "_throughput_note": "output_only",
                     "_concurrency_note": (
                         "client_concurrency is the number of requests sent simultaneously. "
@@ -532,14 +677,44 @@ def _run_online(self, inference_fn: Callable) -> dict:
                 "Pass an async coroutine (inference_fn_streaming), "
                 "not a sync wrapper."
             )
-        loop = asyncio.get_event_loop()
-        return loop.run_until_complete(self._run_online_async(inference_fn))
+        return asyncio.run(self._run_online_async(inference_fn))
+
+    async def _warmup_requests(self, async_inference_fn, count: int, label: str) -> None:
+        """
+        Fire `count` dummy requests sequentially before timed measurement.
+
+        Cycles through self.requests if count > len(requests). All results are
+        discarded — purpose is to JIT-compile kernels, allocate CUDA graphs,
+        prime the KV cache, and let the engine reach steady-state schedules
+        before the timed phase. Without this, the first few timed requests on
+        cold engines inflate p99 by hundreds of milliseconds.
+
+        Exceptions during warmup are logged and swallowed; warmup failures
+        must never abort the timed run.
+        """
+        if count <= 0 or not self.requests:
+            return
+        tqdm.write(
+            f"[{label} warmup] firing {count} dummy requests "
+            "(results discarded — engine JIT/cache warm-up)"
+        )
+        for i in range(count):
+            req = self.requests[i % len(self.requests)]
+            try:
+                await async_inference_fn(req)
+            except Exception as e:
+                tqdm.write(f"[{label} warmup] request {i} failed (ignored): {e}")
 
     async def _run_online_async(self, async_inference_fn) -> dict:
         """
         Async implementation of the online scenario.
         Generates Poisson arrival times upfront, then fires all requests
         concurrently via asyncio.gather so the engine sees real concurrent load.
+
+        A warmup phase fires `online_warmup_requests` dummy requests
+        sequentially before the QPS sweep. Their latencies are not recorded
+        in `results_by_qps`. This prevents cold-engine TTFT spikes from
+        inflating p99 at the first QPS level.
         """
         loop = asyncio.get_event_loop()
         sla_ms = self.suite["online_sla_ttft_ms"]
@@ -547,6 +722,10 @@ async def _run_online_async(self, async_inference_fn) -> dict:
         all_samples: list[SampleRecord] = []
         max_valid_qps = 0.0
 
+        await self._warmup_requests(
+            async_inference_fn, self.online_warmup_requests, "online"
+        )
+
         for target_qps in self.suite["online_qps_levels"]:
             print(f"[online] target_qps={target_qps}")
             run_ttfts: list[list[float]] = []
@@ -607,6 +786,15 @@ async def send_request(req: InferenceRequest, t_arrival: float) -> InferenceResu
             tpot_p90 = float(np.percentile(all_tpots, 90)) if all_tpots else 0
             tpot_p99 = float(np.percentile(all_tpots, 99)) if all_tpots else 0
 
+            # Per-run p99s, used to surface inter-run TTFT variability.
+            # We compute each run's p99 independently; the scenario's overall
+            # `ttft_ms_p99` (above) is computed by pooling all per-request
+            # TTFTs, which is the headline number, while this CV captures
+            # whether that number is reproducible across `num_runs`.
+            ttft_p99_per_run = [
+                float(np.percentile(run, 99)) for run in run_ttfts if run
+            ]
+
             sla_met = ttft_p99 < sla_ms
             if sla_met:
                 max_valid_qps = target_qps
@@ -626,6 +814,8 @@ async def send_request(req: InferenceRequest, t_arrival: float) -> InferenceResu
                 "tpot_ms_p99": round(tpot_p99, 2),
                 "elapsed_seconds_median": round(float(np.median(run_elapsed_times)), 1),
                 "sla_met": sla_met,
+                "ttft_ms_p99_reliability":
+                    _reliability_block(ttft_p99_per_run, decimals=2),
             })
 
         self._write_samples(all_samples)
@@ -665,6 +855,11 @@ async def _run_burst_async(self, async_inference_fn) -> dict:
             sla_met_during_burst    — bool: p99 TTFT during burst < online_sla_ttft_ms
             burst_degradation_ratio — burst_ttft_p99 / steady_ttft_p99 (higher = worse)
             results_by_cycle        — per-cycle breakdown
+
+        A warmup phase fires `burst_warmup_requests` dummy requests
+        sequentially before the first cycle. Their latencies are excluded
+        from steady/burst windows so the first cycle's steady-state
+        measurement is not contaminated by cold-engine TTFT spikes.
         """
         loop = asyncio.get_event_loop()
         sla_ms = self.suite["online_sla_ttft_ms"]
@@ -674,6 +869,10 @@ async def _run_burst_async(self, async_inference_fn) -> dict:
         steady_dur = self.suite["burst_interval_seconds"]
         num_runs = self.suite.get("num_runs", 3)
 
+        await self._warmup_requests(
+            async_inference_fn, self.burst_warmup_requests, "burst"
+        )
+
         all_steady_ttfts: list[float] = []
         all_burst_ttfts: list[float] = []
         results_by_cycle = []
@@ -681,7 +880,16 @@ async def _run_burst_async(self, async_inference_fn) -> dict:
         all_samples: list[SampleRecord] = []
 
         async def fire_window(qps: float, duration_secs: float, label: str):
-            """Fire requests at Poisson QPS for duration_secs. Returns list of InferenceResult."""
+            """
+            Fire requests at Poisson QPS for duration_secs.
+
+            Returns
+                results       : list[InferenceResult] in arrival order
+                elapsed       : wall-clock seconds the window took
+                arrival_times : list[float] — each request's intended arrival
+                                relative to window start (parallel to results).
+                                Used to compute post-burst recovery_time_seconds.
+            """
             n_expected = max(1, int(qps * duration_secs * 1.5))
             requests_pool = (self.requests * ((n_expected // len(self.requests)) + 2))[:n_expected]
 
@@ -690,7 +898,7 @@ async def fire_window(qps: float, duration_secs: float, label: str):
 
             pairs = [(req, t) for req, t in zip(requests_pool, arrival_times) if t < duration_secs]
             if not pairs:
-                return [], 0.0
+                return [], 0.0, []
 
             t_start = loop.time()
 
@@ -702,20 +910,38 @@ async def send(req, t_arrival):
 
             results = list(await asyncio.gather(*[send(req, t) for req, t in pairs]))
             elapsed = loop.time() - t_start
-            return results, elapsed
+            window_arrivals = [t for (_, t) in pairs]
+            return results, elapsed, window_arrivals
+
+        # Each cycle's per-request data, captured so we can compute
+        # recovery_time_seconds in a single post-processing pass after
+        # all cycles complete.
+        cycle_data: list[dict] = []
 
         for cycle_idx in range(num_runs):
             tqdm.write(f"[burst] cycle {cycle_idx + 1}/{num_runs} — steady({steady_qps} qps)...")
 
-            steady_results, steady_elapsed = await fire_window(steady_qps, steady_dur, "steady")
-            steady_ttfts = [r.first_token_time_ms for r in steady_results
-                            if r.success and r.first_token_time_ms is not None]
+            steady_results, steady_elapsed, steady_arrivals = await fire_window(
+                steady_qps, steady_dur, "steady"
+            )
+            steady_ttfts_pairs = [
+                (a, r.first_token_time_ms)
+                for r, a in zip(steady_results, steady_arrivals)
+                if r.success and r.first_token_time_ms is not None
+            ]
+            steady_ttfts = [v for _, v in steady_ttfts_pairs]
 
             tqdm.write(f"[burst] cycle {cycle_idx + 1}/{num_runs} — burst({burst_qps} qps)...")
 
-            burst_results, burst_elapsed = await fire_window(burst_qps, burst_dur, "burst")
-            burst_ttfts = [r.first_token_time_ms for r in burst_results
-                           if r.success and r.first_token_time_ms is not None]
+            burst_results, burst_elapsed, burst_arrivals = await fire_window(
+                burst_qps, burst_dur, "burst"
+            )
+            burst_ttfts_pairs = [
+                (a, r.first_token_time_ms)
+                for r, a in zip(burst_results, burst_arrivals)
+                if r.success and r.first_token_time_ms is not None
+            ]
+            burst_ttfts = [v for _, v in burst_ttfts_pairs]
 
             all_steady_ttfts.extend(steady_ttfts)
             all_burst_ttfts.extend(burst_ttfts)
@@ -723,6 +949,11 @@ async def send(req, t_arrival):
             cycle_steady_p99 = float(np.percentile(steady_ttfts, 99)) if steady_ttfts else None
             cycle_burst_p99  = float(np.percentile(burst_ttfts, 99)) if burst_ttfts else None
 
+            cycle_data.append({
+                "steady_pairs": steady_ttfts_pairs,
+                "burst_pairs":  burst_ttfts_pairs,
+            })
+
             results_by_cycle.append({
                 "cycle": cycle_idx + 1,
                 "steady_requests": len(steady_ttfts),
@@ -748,6 +979,39 @@ async def send(req, t_arrival):
         sla_met_during_burst = (burst_p99 < sla_ms) if burst_p99 is not None else False
         degradation = round(burst_p99 / steady_p99, 3) if (burst_p99 and steady_p99) else None
 
+        # ── Recovery time after burst ─────────────────────────────────────────
+        # Definition: seconds elapsed within a post-burst steady window before
+        # the rolling p99 TTFT drops below 1.5× the long-term steady baseline.
+        #
+        # Implementation: the loop above runs `steady → burst` per cycle, so
+        # cycle (i+1)'s steady window is the post-burst recovery window for
+        # cycle i's burst. We compute one recovery time per cycle that has a
+        # successor steady window, then emit the median (more robust than
+        # mean to a single outlier cycle).
+        recovery_baseline_p99 = steady_p99  # long-term, post-warmup baseline
+        cycle_recovery_times: list[float] = []
+        if recovery_baseline_p99 and recovery_baseline_p99 > 0:
+            threshold = 1.5 * recovery_baseline_p99
+            for i in range(len(cycle_data) - 1):
+                post = cycle_data[i + 1]["steady_pairs"]
+                if not post:
+                    continue
+                arrivals = [a for a, _ in post]
+                ttfts    = [t for _, t in post]
+                rec = _compute_recovery_time(
+                    arrivals, ttfts,
+                    threshold_ms=threshold,
+                    window_s=min(3.0, steady_dur / 2),
+                    min_samples=5,
+                )
+                if rec is not None:
+                    cycle_recovery_times.append(rec)
+
+        recovery_time_seconds = (
+            round(float(np.median(cycle_recovery_times)), 2)
+            if cycle_recovery_times else None
+        )
+
         sla_icon = "✓" if sla_met_during_burst else "✗"
         chip_str = f"  ({self.chip_count} chips)" if self.chip_count > 1 else ""
         tqdm.write(
@@ -772,6 +1036,15 @@ async def send(req, t_arrival):
             "burst_ttft_p99_ms":  round(burst_p99, 2) if burst_p99 else None,
             "sla_met_during_burst": sla_met_during_burst,
             "burst_degradation_ratio": degradation,
+            "recovery_time_seconds": recovery_time_seconds,
+            "recovery_time_seconds_per_cycle": [
+                round(v, 2) for v in cycle_recovery_times
+            ] if cycle_recovery_times else [],
+            "_recovery_definition": (
+                "Median seconds within the post-burst steady window before "
+                "rolling TTFT p99 drops below 1.5x the long-term steady baseline. "
+                "Lower is better; None means it never recovered within the window."
+            ),
             "results_by_cycle": results_by_cycle,
         }}
 
@@ -781,8 +1054,7 @@ def _run_burst(self, inference_fn: Callable) -> dict:
             raise TypeError(
                 "_run_burst requires an async inference_fn(request: InferenceRequest) -> InferenceResult."
             )
-        loop = asyncio.get_event_loop()
-        return loop.run_until_complete(self._run_burst_async(inference_fn))
+        return asyncio.run(self._run_burst_async(inference_fn))
 
     # ------------------------------------------------------------------
     # Interactive scenario
@@ -793,11 +1065,15 @@ async def _run_interactive_async(self, async_inference_fn) -> dict:
         Send one request at a time, waiting for completion before sending the next.
         Measures single-request latency in isolation (no queueing pressure).
         Uses the same async engine as online to ensure consistent TTFT measurement.
+
+        Per-run TTFT p99s are captured so the result emits an inter-run
+        reliability block alongside the pooled metrics.
         """
         all_ttfts: list[float] = []
         all_tpots: list[float] = []
         all_samples: list[SampleRecord] = []
         run_elapsed_times: list[float] = []
+        ttft_p99_per_run: list[float] = []
 
         total_runs = self.warmup_runs + self.suite["num_runs"]
 
@@ -848,6 +1124,8 @@ async def _run_interactive_async(self, async_inference_fn) -> dict:
             all_ttfts.extend(run_ttfts)
             all_tpots.extend(run_tpots)
             run_elapsed_times.append(run_elapsed)
+            if run_ttfts:
+                ttft_p99_per_run.append(float(np.percentile(run_ttfts, 99)))
 
             if run_ttfts:
                 tqdm.write(
@@ -869,6 +1147,8 @@ async def _run_interactive_async(self, async_inference_fn) -> dict:
             "tpot_ms_p99": round(float(np.percentile(all_tpots, 99)), 2) if all_tpots else None,
             "peak_memory_gb": None,
             "elapsed_seconds_median": round(float(np.median(run_elapsed_times)), 1) if run_elapsed_times else None,
+            "ttft_ms_p99_reliability":
+                _reliability_block(ttft_p99_per_run, decimals=2),
         }}
 
     # ------------------------------------------------------------------
diff --git a/loadgen/tests/__init__.py b/loadgen/tests/__init__.py
new file mode 100644
index 00000000..e69de29b
diff --git a/loadgen/tests/test_reliability.py b/loadgen/tests/test_reliability.py
new file mode 100644
index 00000000..c2d4ea1b
--- /dev/null
+++ b/loadgen/tests/test_reliability.py
@@ -0,0 +1,243 @@
+"""
+Tests for the reliability blocks emitted by each loadgen scenario.
+
+Locks down:
+- `_cv_pct` / `_stability_label` helpers
+- `_reliability_block` shape contract
+- `_compute_recovery_time` rolling-window logic
+- offline / online / interactive / sustained / burst each emit the new
+  fields with the expected types and a non-None CV when n >= 2
+
+These tests use the same MockInferenceFn pattern as test_warmup.py — a
+real `async def` closure bound to a counter, since loadgen detects
+coroutines via `asyncio.iscoroutinefunction()`.
+"""
+
+from __future__ import annotations
+
+import asyncio
+
+import pytest
+
+from loadgen.loadgen import (
+    AccelMarkLoadGen,
+    _compute_recovery_time,
+    _cv_pct,
+    _reliability_block,
+    _stability_label,
+)
+from loadgen.types import InferenceResult
+
+
+# ── Pure helper tests ─────────────────────────────────────────────────────────
+
+def test_cv_pct_basic():
+    assert _cv_pct([100.0, 100.0, 100.0]) == 0.0
+    cv = _cv_pct([90.0, 100.0, 110.0])
+    assert cv is not None
+    assert 9.0 < cv < 11.0, f"expected CV near 10%, got {cv}"
+
+
+def test_cv_pct_returns_none_for_small_or_invalid_input():
+    assert _cv_pct([]) is None
+    assert _cv_pct([42.0]) is None
+    assert _cv_pct([0.0, 0.0, 0.0]) is None  # mean=0, undefined CV
+
+
+def test_stability_labels():
+    # Boundaries: ≤3% stable, ≤8% noisy, >8% high-variance.
+    assert _stability_label(0.5) == "stable"
+    assert _stability_label(3.0) == "stable"          # inclusive boundary
+    assert _stability_label(3.01) == "noisy"
+    assert _stability_label(8.0) == "noisy"           # inclusive boundary
+    assert _stability_label(8.01) == "high-variance"
+    assert _stability_label(20.0) == "high-variance"
+    assert _stability_label(None) is None
+
+
+def test_reliability_block_shape():
+    block = _reliability_block([100.0, 102.0, 98.0], decimals=1)
+    assert set(block.keys()) == {"n", "mean", "std", "cv_pct", "stability", "runs"}
+    assert block["n"] == 3
+    assert block["mean"] == 100.0
+    assert block["runs"] == [100.0, 102.0, 98.0]
+    assert block["stability"] == "stable"
+
+
+def test_reliability_block_empty_input_returns_empty_dict():
+    """Frontend gates on the block being non-empty; never None."""
+    assert _reliability_block([]) == {}
+
+
+# ── Recovery-time tests ──────────────────────────────────────────────────────
+
+def test_recovery_time_finds_the_first_clean_window():
+    """Build a synthetic post-burst window where the first 5 seconds are
+    elevated and everything after is clean. Recovery must land around 5s."""
+    arrivals = [i * 0.5 for i in range(40)]   # 20 seconds of arrivals at 2 Hz
+    # Elevated TTFTs first 5 s, then drop to clean values.
+    ttfts = [1500.0 if a < 5.0 else 200.0 for a in arrivals]
+    rec = _compute_recovery_time(arrivals, ttfts, threshold_ms=500.0, window_s=2.0, min_samples=4)
+    assert rec is not None, "expected recovery, got None"
+    assert 4.5 <= rec <= 8.0, f"recovery expected ≈5–8s, got {rec}"
+
+
+def test_recovery_time_returns_none_when_never_recovers():
+    arrivals = [i * 0.5 for i in range(20)]
+    ttfts = [2000.0] * 20  # always above any sane threshold
+    assert _compute_recovery_time(arrivals, ttfts, threshold_ms=500.0) is None
+
+
+def test_recovery_time_returns_none_when_too_few_samples():
+    assert _compute_recovery_time([], [], threshold_ms=500.0) is None
+    assert _compute_recovery_time([1.0, 2.0], [100.0, 100.0],
+                                  threshold_ms=500.0, min_samples=5) is None
+
+
+# ── Scenario integration tests ───────────────────────────────────────────────
+
+def _make_requests(n: int):
+    from loadgen.loadgen import InferenceRequest
+    return [
+        InferenceRequest(prompt=f"p{i}", request_id=i, input_tokens=10, max_tokens=20)
+        for i in range(n)
+    ]
+
+
+def _async_fn(ttft_ms: float = 100.0):
+    """Build a real `async def` returning a constant InferenceResult."""
+    async def fn(request) -> InferenceResult:
+        await asyncio.sleep(0)
+        return InferenceResult(
+            first_token_time_ms=ttft_ms,
+            total_time_ms=ttft_ms * 2,
+            output_tokens=20,
+            input_tokens=10,
+            success=True,
+        )
+    return fn
+
+
+def _sync_offline_fn(ttft_ms: float = 100.0):
+    """Sync inference_fn used for offline scenario — receives list of requests."""
+    def fn(reqs):
+        return [
+            InferenceResult(
+                first_token_time_ms=ttft_ms,
+                total_time_ms=ttft_ms * 2,
+                output_tokens=20,
+                input_tokens=10,
+                success=True,
+            )
+            for _ in reqs
+        ]
+    return fn
+
+
+def test_offline_emits_throughput_reliability(tmp_path):
+    suite = {
+        "concurrency_levels": [4],
+        "num_runs": 3,
+        "warmup_runs": 0,
+        "request_count": 8,
+        "input_tokens": 10,
+    }
+    requests = _make_requests(8)
+    gen = AccelMarkLoadGen(suite, requests, "offline", str(tmp_path))
+    result = gen.run(_sync_offline_fn())
+
+    cc_results = result["offline"]["results_by_concurrency"]
+    assert cc_results, "offline scenario produced no results"
+    rel = cc_results[0].get("throughput_tokens_per_sec_reliability")
+    assert rel, "offline scenario did not emit reliability block"
+    assert rel["n"] == 3
+    assert rel["cv_pct"] is not None
+    assert rel["stability"] in {"stable", "noisy", "high-variance"}
+    assert len(rel["runs"]) == 3
+
+
+def test_online_emits_ttft_p99_reliability(tmp_path):
+    suite = {
+        "num_runs": 2,
+        "online_qps_levels": [2.0],
+        "online_sla_ttft_ms": 1000,
+        "online_request_count": 6,
+        "online_warmup_requests": 0,
+        "input_tokens": 10,
+    }
+    requests = _make_requests(6)
+    gen = AccelMarkLoadGen(suite, requests, "online", str(tmp_path))
+    result = gen.run(_async_fn(ttft_ms=100.0))
+
+    qps_results = result["online"]["results_by_qps"]
+    assert qps_results, "online scenario produced no results"
+    rel = qps_results[0].get("ttft_ms_p99_reliability")
+    assert rel, "online scenario did not emit reliability block"
+    assert rel["n"] == 2
+    # With constant TTFT the CV should be exactly 0.
+    assert rel["cv_pct"] == 0.0
+    assert rel["stability"] == "stable"
+
+
+def test_interactive_emits_ttft_p99_reliability(tmp_path):
+    suite = {
+        "num_runs": 2,
+        "interactive_warmup_runs": 0,
+        "interactive_request_count": 4,
+        "input_tokens": 10,
+    }
+    requests = _make_requests(4)
+    gen = AccelMarkLoadGen(suite, requests, "interactive", str(tmp_path))
+    result = gen.run(_async_fn(ttft_ms=120.0))
+
+    inter = result["interactive"]
+    rel = inter.get("ttft_ms_p99_reliability")
+    assert rel, "interactive scenario did not emit reliability block"
+    assert rel["n"] == 2
+    assert rel["stability"] == "stable"
+
+
+def test_sustained_emits_throughput_post_warmup_reliability(tmp_path):
+    """Run a tiny sustained scenario — long enough to produce ≥2 sample
+    intervals so CV is computable."""
+    suite = {
+        "sustained_concurrency": 2,
+        "duration_minutes": 4 / 60,        # 4 seconds total
+        "sample_interval_seconds": 1.0,
+        "warmup_minutes": 1 / 60,          # 1-second warmup
+        "input_tokens": 10,
+    }
+    requests = _make_requests(20)
+    gen = AccelMarkLoadGen(suite, requests, "sustained", str(tmp_path))
+    result = gen.run(_async_fn(ttft_ms=30.0))
+
+    rel = result["sustained"].get("throughput_post_warmup_reliability")
+    assert isinstance(rel, dict), "sustained scenario did not emit reliability block"
+    # cv_pct may be None if not enough post-warmup samples landed; we only
+    # require the field exists. When n >= 2 the stability must be set.
+    if rel.get("n", 0) >= 2:
+        assert rel["stability"] in {"stable", "noisy", "high-variance"}
+
+
+def test_burst_emits_recovery_time_seconds(tmp_path):
+    """Burst with constant low TTFT should report a finite (small)
+    recovery_time and a list (possibly empty) per-cycle field."""
+    suite = {
+        "num_runs": 2,
+        "online_sla_ttft_ms": 1000,
+        "online_request_count": 6,
+        "burst_warmup_requests": 0,
+        "burst_steady_qps": 2.0,
+        "burst_peak_qps": 4.0,
+        "burst_duration_seconds": 0.5,
+        "burst_interval_seconds": 0.5,
+        "input_tokens": 10,
+    }
+    requests = _make_requests(6)
+    gen = AccelMarkLoadGen(suite, requests, "burst", str(tmp_path))
+    result = gen.run(_async_fn(ttft_ms=100.0))
+
+    burst = result["burst"]
+    assert "recovery_time_seconds" in burst
+    assert "recovery_time_seconds_per_cycle" in burst
+    assert isinstance(burst["recovery_time_seconds_per_cycle"], list)
diff --git a/loadgen/tests/test_warmup.py b/loadgen/tests/test_warmup.py
new file mode 100644
index 00000000..ea625986
--- /dev/null
+++ b/loadgen/tests/test_warmup.py
@@ -0,0 +1,254 @@
+"""
+Tests for the warmup phase in online and burst scenarios.
+
+These scenarios used to read `online_warmup_runs` from suite.json but
+silently ignored the value — every reported p99 was contaminated by
+cold-engine TTFT spikes. This test suite locks down the fix so future
+refactors can't reintroduce the bug.
+
+What is verified:
+- Warmup requests are fired in `online` and `burst` BEFORE the timed phase
+- Warmup latencies are NOT counted in the returned distribution
+- A counter on the mock inference_fn confirms the exact request budget
+- Warmup is a no-op when the parameter is 0 (back-compat)
+- An exception during warmup does not abort the timed phase
+"""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Optional
+
+import pytest
+
+from loadgen.loadgen import AccelMarkLoadGen
+from loadgen.types import InferenceResult
+
+
+# ── Fixtures ──────────────────────────────────────────────────────────────────
+
+def _make_requests(n: int):
+    """Build n minimal InferenceRequest-like objects using the same shim
+    loadgen.py falls back to when benchmark_runner is not importable."""
+    from loadgen.loadgen import InferenceRequest
+    return [
+        InferenceRequest(
+            prompt=f"prompt {i}",
+            request_id=i,
+            input_tokens=10,
+            max_tokens=20,
+        )
+        for i in range(n)
+    ]
+
+
+def _online_suite(qps_levels=(2.0,), warmup_requests: int = 5, num_runs: int = 1):
+    return {
+        "num_runs": num_runs,
+        "online_qps_levels": list(qps_levels),
+        "online_sla_ttft_ms": 1000,
+        "online_request_count": 8,
+        "online_warmup_requests": warmup_requests,
+        "input_tokens": 10,
+    }
+
+
+def _burst_suite(warmup_requests: int = 5):
+    return {
+        "num_runs": 1,
+        "online_sla_ttft_ms": 1000,
+        "online_request_count": 6,
+        "burst_steady_qps": 2.0,
+        "burst_peak_qps": 4.0,
+        "burst_duration_seconds": 0.3,
+        "burst_interval_seconds": 0.3,
+        "burst_warmup_requests": warmup_requests,
+        "input_tokens": 10,
+    }
+
+
+class MockInferenceFn:
+    """Counts every call and exposes an async callable as `.fn` for loadgen.
+
+    The fast warmup latency vs slow timed latency makes it trivial to
+    assert that warmup requests are excluded from the distribution: if
+    warmup latencies leaked into results, p50/p99 would collapse to
+    the fast value.
+
+    Note: loadgen uses `asyncio.iscoroutinefunction()` to detect async
+    inference_fn, which returns False for a class with `async __call__`.
+    So we expose `self.fn` as a real `async def` closure bound to this
+    instance's state.
+    """
+
+    def __init__(self, *, warmup_ttft_ms: float = 1.0, timed_ttft_ms: float = 100.0,
+                 fail_first_n: int = 0):
+        self.call_count = 0
+        self.warmup_ttft_ms = warmup_ttft_ms
+        self.timed_ttft_ms = timed_ttft_ms
+        self.fail_first_n = fail_first_n
+        self.warmup_budget: Optional[int] = None
+
+        state = self  # closure capture
+
+        async def _fn(request) -> InferenceResult:
+            idx = state.call_count
+            state.call_count += 1
+            if idx < state.fail_first_n:
+                raise RuntimeError(f"simulated failure on warmup request {idx}")
+            ttft = (
+                state.warmup_ttft_ms
+                if state.warmup_budget is not None and idx < state.warmup_budget
+                else state.timed_ttft_ms
+            )
+            await asyncio.sleep(0)  # yield control
+            return InferenceResult(
+                first_token_time_ms=ttft,
+                total_time_ms=ttft * 2,
+                output_tokens=20,
+                input_tokens=10,
+                success=True,
+            )
+
+        self.fn = _fn
+
+    def set_warmup_budget(self, n: int) -> None:
+        """Tell the mock how many of the next calls count as warmup."""
+        self.warmup_budget = n
+
+
+# ── online warmup ─────────────────────────────────────────────────────────────
+
+def test_online_warmup_fires_configured_count(tmp_path):
+    """Online scenario must fire exactly `online_warmup_requests` warmup calls."""
+    suite = _online_suite(qps_levels=(2.0,), warmup_requests=5)
+    requests = _make_requests(8)
+    gen = AccelMarkLoadGen(suite, requests, "online", str(tmp_path))
+
+    fn = MockInferenceFn()
+    fn.set_warmup_budget(5)
+    gen.run(fn.fn)
+
+    # warmup (5) + 8 requests × 1 QPS × 1 run = 13 calls minimum
+    assert fn.call_count >= 13, (
+        f"expected at least 13 inference_fn calls (5 warmup + 8 timed), "
+        f"got {fn.call_count}"
+    )
+
+
+def test_online_warmup_latencies_excluded_from_p99(tmp_path):
+    """If warmup latencies leaked into the recorded distribution, p99 would
+    collapse to the fast warmup value. Verify it stays at the timed value."""
+    suite = _online_suite(qps_levels=(2.0,), warmup_requests=5)
+    requests = _make_requests(8)
+    gen = AccelMarkLoadGen(suite, requests, "online", str(tmp_path))
+
+    fn = MockInferenceFn(warmup_ttft_ms=1.0, timed_ttft_ms=100.0)
+    fn.set_warmup_budget(5)
+    result = gen.run(fn.fn)
+
+    qps_results = result["online"]["results_by_qps"]
+    assert qps_results, "expected at least one QPS level result"
+    p50 = qps_results[0]["ttft_ms_p50"]
+    p99 = qps_results[0]["ttft_ms_p99"]
+
+    # If warmup leaked in, p50 would be near 1.0 ms. With warmup excluded,
+    # every recorded request returns 100 ms, so all percentiles snap there.
+    assert abs(p50 - 100.0) < 0.5, f"p50 contaminated by warmup: {p50}"
+    assert abs(p99 - 100.0) < 0.5, f"p99 contaminated by warmup: {p99}"
+
+
+def test_online_warmup_zero_is_noop(tmp_path):
+    """Backward compat: setting online_warmup_requests=0 must skip warmup."""
+    suite = _online_suite(qps_levels=(2.0,), warmup_requests=0)
+    requests = _make_requests(6)
+    gen = AccelMarkLoadGen(suite, requests, "online", str(tmp_path))
+
+    fn = MockInferenceFn()
+    gen.run(fn.fn)
+
+    # No warmup means exactly 6 calls (1 QPS × 6 requests × 1 run).
+    assert fn.call_count == 6, (
+        f"warmup=0 should fire only timed requests; got {fn.call_count} calls"
+    )
+
+
+def test_online_warmup_failure_does_not_abort_run(tmp_path):
+    """A failing warmup request must be logged and ignored — the timed phase
+    must still execute. Otherwise a flaky engine could prevent any submission."""
+    suite = _online_suite(qps_levels=(2.0,), warmup_requests=3)
+    requests = _make_requests(6)
+    gen = AccelMarkLoadGen(suite, requests, "online", str(tmp_path))
+
+    # Fail the first 2 warmup requests; the 3rd warmup + all timed must run.
+    fn = MockInferenceFn(fail_first_n=2)
+    fn.set_warmup_budget(3)
+    result = gen.run(fn.fn)
+
+    qps_results = result["online"]["results_by_qps"]
+    assert qps_results, "timed phase did not run despite warmup failures"
+    assert fn.call_count >= 3 + 6  # 3 warmup attempts + 6 timed
+
+
+# ── burst warmup ──────────────────────────────────────────────────────────────
+
+def test_burst_warmup_fires_configured_count(tmp_path):
+    suite = _burst_suite(warmup_requests=4)
+    requests = _make_requests(6)
+    gen = AccelMarkLoadGen(suite, requests, "burst", str(tmp_path))
+
+    fn = MockInferenceFn()
+    fn.set_warmup_budget(4)
+    gen.run(fn.fn)
+
+    # At least 4 warmup calls must have fired before the timed cycles.
+    assert fn.call_count >= 4, (
+        f"burst warmup did not fire enough requests: {fn.call_count}"
+    )
+
+
+def test_burst_warmup_zero_is_noop(tmp_path):
+    """Suites that omit burst_warmup_requests entirely default to 10; setting
+    it to 0 must skip warmup."""
+    suite = _burst_suite(warmup_requests=0)
+    requests = _make_requests(6)
+    gen = AccelMarkLoadGen(suite, requests, "burst", str(tmp_path))
+
+    fn = MockInferenceFn()
+    n_before = fn.call_count
+    gen.run(fn.fn)
+    # No assertion on exact count (timed cycles depend on Poisson timing),
+    # but we can assert the mock saw at least 1 timed call.
+    assert fn.call_count > n_before
+
+
+# ── default values ────────────────────────────────────────────────────────────
+
+def test_online_warmup_default_is_ten(tmp_path):
+    """Suite without online_warmup_requests should get a sensible default."""
+    suite = {
+        "num_runs": 1,
+        "online_qps_levels": [2.0],
+        "online_sla_ttft_ms": 1000,
+        "online_request_count": 4,
+        "input_tokens": 10,
+    }
+    requests = _make_requests(4)
+    gen = AccelMarkLoadGen(suite, requests, "online", str(tmp_path))
+    assert gen.online_warmup_requests == 10
+
+
+def test_burst_warmup_default_is_ten(tmp_path):
+    suite = {
+        "num_runs": 1,
+        "online_sla_ttft_ms": 1000,
+        "online_request_count": 4,
+        "burst_steady_qps": 2.0,
+        "burst_peak_qps": 4.0,
+        "burst_duration_seconds": 0.3,
+        "burst_interval_seconds": 0.3,
+        "input_tokens": 10,
+    }
+    requests = _make_requests(4)
+    gen = AccelMarkLoadGen(suite, requests, "burst", str(tmp_path))
+    assert gen.burst_warmup_requests == 10
diff --git a/results/community/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_sglang_c43a8309_958afbbd/result.json b/results/community/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_sglang_c43a8309_958afbbd/result.json
index cf46b029..adcbafe3 100644
--- a/results/community/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_sglang_c43a8309_958afbbd/result.json
+++ b/results/community/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_sglang_c43a8309_958afbbd/result.json
@@ -496,7 +496,44 @@
       "sustained_throughput_tokens_per_sec": 562.5,
       "throttle_ratio": 0.966,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -0.3
+      "ttft_p99_drift_ms": -0.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 562.5,
+        "std": 4.9,
+        "cv_pct": 0.86,
+        "stability": "stable",
+        "runs": [
+          566.2,
+          560.8,
+          565.3,
+          561.4,
+          561.9,
+          570.0,
+          558.3,
+          563.3,
+          552.7,
+          569.3,
+          558.9,
+          568.4,
+          557.2,
+          565.5,
+          554.1,
+          563.7,
+          563.3,
+          565.0,
+          563.3,
+          564.7,
+          564.3,
+          562.3,
+          569.5,
+          550.8,
+          562.5,
+          561.7,
+          566.5,
+          558.2
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/community/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_sglang_c43a8309_99c43b97/result.json b/results/community/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_sglang_c43a8309_99c43b97/result.json
index fb4ac5fd..e110ab18 100644
--- a/results/community/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_sglang_c43a8309_99c43b97/result.json
+++ b/results/community/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_sglang_c43a8309_99c43b97/result.json
@@ -436,7 +436,44 @@
       "sustained_throughput_tokens_per_sec": 54.9,
       "throttle_ratio": 0.666,
       "throttle_onset_minute": 11,
-      "ttft_p99_drift_ms": 66
+      "ttft_p99_drift_ms": 66,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 54.9,
+        "std": 4.9,
+        "cv_pct": 8.95,
+        "stability": "high-variance",
+        "runs": [
+          56.2,
+          56.3,
+          56.2,
+          56.3,
+          56.3,
+          56.2,
+          56.2,
+          56.2,
+          56.3,
+          37.5,
+          56.3,
+          56.2,
+          56.3,
+          56.2,
+          56.3,
+          56.2,
+          56.3,
+          56.2,
+          56.3,
+          56.2,
+          37.5,
+          56.3,
+          56.2,
+          56.2,
+          56.3,
+          56.3,
+          56.2,
+          56.2
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/community/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_sglang_c43a8309_435424a8/result.json b/results/community/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_sglang_c43a8309_435424a8/result.json
index 29bb5796..54c9f403 100644
--- a/results/community/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_sglang_c43a8309_435424a8/result.json
+++ b/results/community/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_sglang_c43a8309_435424a8/result.json
@@ -327,7 +327,30 @@
       "sustained_throughput_tokens_per_sec": 7095.4,
       "throttle_ratio": 0.92,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -4707.7
+      "ttft_p99_drift_ms": -4707.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 7095.4,
+        "std": 147.1,
+        "cv_pct": 2.07,
+        "stability": "stable",
+        "runs": [
+          6616.4,
+          7181.7,
+          7188.9,
+          7110.2,
+          7106.5,
+          7144.4,
+          7158.3,
+          7020.0,
+          7160.7,
+          7183.9,
+          7180.8,
+          7127.9,
+          7111.1,
+          7044.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/result.json b/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/result.json
index 1652c2b9..cfe652b4 100644
--- a/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/result.json
+++ b/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/result.json
@@ -573,7 +573,44 @@
       "sustained_throughput_tokens_per_sec": 268.0,
       "throttle_ratio": 0.868,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -21.4
+      "ttft_p99_drift_ms": -21.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 268.0,
+        "std": 8.9,
+        "cv_pct": 3.3,
+        "stability": "noisy",
+        "runs": [
+          251.1,
+          271.0,
+          265.9,
+          264.0,
+          276.5,
+          269.5,
+          247.8,
+          277.0,
+          283.2,
+          262.3,
+          271.0,
+          266.3,
+          268.9,
+          252.2,
+          267.2,
+          263.6,
+          271.2,
+          285.4,
+          270.5,
+          266.6,
+          270.5,
+          277.3,
+          266.5,
+          259.0,
+          279.0,
+          268.9,
+          271.4,
+          260.1
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/sustained/result.json b/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/sustained/result.json
index 6aa62562..8ca9962f 100644
--- a/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/sustained/result.json
+++ b/results/verified/huawei_ascend_910b2x1_suite_A_ascend_vllm_ascend_d4aa9fda_a2777c30/sustained/result.json
@@ -475,7 +475,44 @@
       "sustained_throughput_tokens_per_sec": 268.0,
       "throttle_ratio": 0.868,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -21.4
+      "ttft_p99_drift_ms": -21.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 268.0,
+        "std": 8.9,
+        "cv_pct": 3.3,
+        "stability": "noisy",
+        "runs": [
+          251.1,
+          271.0,
+          265.9,
+          264.0,
+          276.5,
+          269.5,
+          247.8,
+          277.0,
+          283.2,
+          262.3,
+          271.0,
+          266.3,
+          268.9,
+          252.2,
+          267.2,
+          263.6,
+          271.2,
+          285.4,
+          270.5,
+          266.6,
+          270.5,
+          277.3,
+          266.5,
+          259.0,
+          279.0,
+          268.9,
+          271.4,
+          260.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/result.json b/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/result.json
index 3eef1882..442ac69f 100644
--- a/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/result.json
+++ b/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/result.json
@@ -517,7 +517,44 @@
       "sustained_throughput_tokens_per_sec": 53.2,
       "throttle_ratio": 0.733,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -20614.1
+      "ttft_p99_drift_ms": -20614.1,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 53.2,
+        "std": 5.6,
+        "cv_pct": 10.6,
+        "stability": "high-variance",
+        "runs": [
+          46.9,
+          59.7,
+          51.2,
+          51.2,
+          55.5,
+          46.9,
+          64.0,
+          46.9,
+          59.7,
+          46.9,
+          55.5,
+          55.5,
+          46.9,
+          59.8,
+          46.9,
+          59.7,
+          46.9,
+          59.7,
+          51.2,
+          51.2,
+          55.4,
+          46.9,
+          64.0,
+          46.9,
+          55.5,
+          51.2,
+          55.5,
+          51.2
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/sustained/result.json b/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/sustained/result.json
index f8184d71..c4af0d0a 100644
--- a/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/sustained/result.json
+++ b/results/verified/huawei_ascend_910b2x1_suite_D_ascend_vllm_ascend_d4aa9fda_a3547ba9/sustained/result.json
@@ -475,7 +475,44 @@
       "sustained_throughput_tokens_per_sec": 53.2,
       "throttle_ratio": 0.733,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -20614.1
+      "ttft_p99_drift_ms": -20614.1,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 53.2,
+        "std": 5.6,
+        "cv_pct": 10.6,
+        "stability": "high-variance",
+        "runs": [
+          46.9,
+          59.7,
+          51.2,
+          51.2,
+          55.5,
+          46.9,
+          64.0,
+          46.9,
+          59.7,
+          46.9,
+          55.5,
+          55.5,
+          46.9,
+          59.8,
+          46.9,
+          59.7,
+          46.9,
+          59.7,
+          51.2,
+          51.2,
+          55.4,
+          46.9,
+          64.0,
+          46.9,
+          55.5,
+          51.2,
+          55.5,
+          51.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/result.json b/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/result.json
index 67225a10..9e65b7bf 100644
--- a/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/result.json
+++ b/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/result.json
@@ -409,7 +409,30 @@
       "sustained_throughput_tokens_per_sec": 1238.9,
       "throttle_ratio": 0.883,
       "throttle_onset_minute": 6.0,
-      "ttft_p99_drift_ms": -302.5
+      "ttft_p99_drift_ms": -302.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1238.9,
+        "std": 47.1,
+        "cv_pct": 3.8,
+        "stability": "noisy",
+        "runs": [
+          1230.3,
+          1302.7,
+          1288.1,
+          1349.2,
+          1277.4,
+          1204.2,
+          1213.4,
+          1211.4,
+          1191.5,
+          1223.0,
+          1239.3,
+          1204.1,
+          1199.3,
+          1210.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/sustained/result.json b/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/sustained/result.json
index 0d8c6baf..113b11e6 100644
--- a/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/sustained/result.json
+++ b/results/verified/huawei_ascend_910b2x1_suite_F_ascend_vllm_ascend_d4aa9fda_bd7d8f87/sustained/result.json
@@ -325,7 +325,30 @@
       "sustained_throughput_tokens_per_sec": 1238.9,
       "throttle_ratio": 0.883,
       "throttle_onset_minute": 6.0,
-      "ttft_p99_drift_ms": -302.5
+      "ttft_p99_drift_ms": -302.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1238.9,
+        "std": 47.1,
+        "cv_pct": 3.8,
+        "stability": "noisy",
+        "runs": [
+          1230.3,
+          1302.7,
+          1288.1,
+          1349.2,
+          1277.4,
+          1204.2,
+          1213.4,
+          1211.4,
+          1191.5,
+          1223.0,
+          1239.3,
+          1204.1,
+          1199.3,
+          1210.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/result.json b/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/result.json
index 3061b81d..1c40ff18 100644
--- a/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/result.json
+++ b/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/result.json
@@ -574,7 +574,44 @@
       "sustained_throughput_tokens_per_sec": 53.2,
       "throttle_ratio": 0.616,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -127.8
+      "ttft_p99_drift_ms": -127.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 53.2,
+        "std": 7.0,
+        "cv_pct": 13.23,
+        "stability": "high-variance",
+        "runs": [
+          56.7,
+          55.0,
+          54.4,
+          54.2,
+          52.3,
+          40.7,
+          62.1,
+          54.1,
+          52.1,
+          63.0,
+          40.6,
+          65.9,
+          41.9,
+          44.9,
+          65.4,
+          50.3,
+          58.2,
+          52.0,
+          49.0,
+          55.4,
+          49.7,
+          51.5,
+          54.8,
+          55.4,
+          63.5,
+          45.1,
+          55.0,
+          45.3
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 151.0,
diff --git a/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/sustained/result.json b/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/sustained/result.json
index 9521952a..1511541c 100644
--- a/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/sustained/result.json
+++ b/results/verified/huawei_ascend_910b2x8_suite_B_ascend_vllm_ascend_d4aa9fda_fcb9725c/sustained/result.json
@@ -475,7 +475,44 @@
       "sustained_throughput_tokens_per_sec": 53.2,
       "throttle_ratio": 0.616,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -127.8
+      "ttft_p99_drift_ms": -127.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 53.2,
+        "std": 7.0,
+        "cv_pct": 13.23,
+        "stability": "high-variance",
+        "runs": [
+          56.7,
+          55.0,
+          54.4,
+          54.2,
+          52.3,
+          40.7,
+          62.1,
+          54.1,
+          52.1,
+          63.0,
+          40.6,
+          65.9,
+          41.9,
+          44.9,
+          65.4,
+          50.3,
+          58.2,
+          52.0,
+          49.0,
+          55.4,
+          49.7,
+          51.5,
+          54.8,
+          55.4,
+          63.5,
+          45.1,
+          55.0,
+          45.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/result.json b/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/result.json
index 1941f8ac..eb69665d 100644
--- a/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/result.json
+++ b/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/result.json
@@ -571,7 +571,44 @@
       "sustained_throughput_tokens_per_sec": 226.6,
       "throttle_ratio": 0.821,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -109.1
+      "ttft_p99_drift_ms": -109.1,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 226.6,
+        "std": 11.3,
+        "cv_pct": 5.01,
+        "stability": "noisy",
+        "runs": [
+          217.7,
+          215.0,
+          227.9,
+          236.8,
+          220.7,
+          231.6,
+          213.2,
+          230.9,
+          230.7,
+          225.5,
+          221.6,
+          244.7,
+          201.0,
+          231.4,
+          239.6,
+          239.7,
+          220.8,
+          225.3,
+          238.2,
+          201.8,
+          239.8,
+          226.4,
+          222.9,
+          219.6,
+          241.1,
+          236.3,
+          213.8,
+          231.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/sustained/result.json b/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/sustained/result.json
index abb80572..9c23b4c0 100644
--- a/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/sustained/result.json
+++ b/results/verified/huawei_ascend_910b2x8_suite_G_ascend_vllm_ascend_d4aa9fda_d726144e/sustained/result.json
@@ -475,7 +475,44 @@
       "sustained_throughput_tokens_per_sec": 226.6,
       "throttle_ratio": 0.821,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -109.1
+      "ttft_p99_drift_ms": -109.1,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 226.6,
+        "std": 11.3,
+        "cv_pct": 5.01,
+        "stability": "noisy",
+        "runs": [
+          217.7,
+          215.0,
+          227.9,
+          236.8,
+          220.7,
+          231.6,
+          213.2,
+          230.9,
+          230.7,
+          225.5,
+          221.6,
+          244.7,
+          201.0,
+          231.4,
+          239.6,
+          239.7,
+          220.8,
+          225.3,
+          238.2,
+          201.8,
+          239.8,
+          226.4,
+          222.9,
+          219.6,
+          241.1,
+          236.3,
+          213.8,
+          231.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/result.json b/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/result.json
index 7b12b9eb..2b038c2a 100644
--- a/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/result.json
+++ b/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/result.json
@@ -621,7 +621,44 @@
       "sustained_throughput_tokens_per_sec": 53.5,
       "throttle_ratio": 0.603,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -14.8
+      "ttft_p99_drift_ms": -14.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 53.5,
+        "std": 7.9,
+        "cv_pct": 14.84,
+        "stability": "high-variance",
+        "runs": [
+          67.8,
+          42.6,
+          54.7,
+          54.2,
+          54.7,
+          64.5,
+          52.3,
+          41.8,
+          48.5,
+          63.0,
+          47.9,
+          60.4,
+          59.1,
+          54.8,
+          48.3,
+          50.3,
+          50.4,
+          54.4,
+          63.6,
+          48.6,
+          56.1,
+          46.2,
+          60.1,
+          40.9,
+          67.1,
+          42.5,
+          60.5,
+          43.5
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 152.7,
diff --git a/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/sustained/result.json b/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/sustained/result.json
index 6bd84a6a..f87c4b33 100644
--- a/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/sustained/result.json
+++ b/results/verified/huawei_ascend_ascend910x16_suite_B_ascend_vllm_ascend_d4aa9fda_635ecf42/sustained/result.json
@@ -522,7 +522,44 @@
       "sustained_throughput_tokens_per_sec": 53.5,
       "throttle_ratio": 0.603,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -14.8
+      "ttft_p99_drift_ms": -14.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 53.5,
+        "std": 7.9,
+        "cv_pct": 14.84,
+        "stability": "high-variance",
+        "runs": [
+          67.8,
+          42.6,
+          54.7,
+          54.2,
+          54.7,
+          64.5,
+          52.3,
+          41.8,
+          48.5,
+          63.0,
+          47.9,
+          60.4,
+          59.1,
+          54.8,
+          48.3,
+          50.3,
+          50.4,
+          54.4,
+          63.6,
+          48.6,
+          56.1,
+          46.2,
+          60.1,
+          40.9,
+          67.1,
+          42.5,
+          60.5,
+          43.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/result.json b/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/result.json
index 6bbc8bb0..284b5456 100644
--- a/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/result.json
+++ b/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/result.json
@@ -618,7 +618,44 @@
       "sustained_throughput_tokens_per_sec": 262.2,
       "throttle_ratio": 0.861,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": -18.4
+      "ttft_p99_drift_ms": -18.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 262.2,
+        "std": 9.5,
+        "cv_pct": 3.62,
+        "stability": "noisy",
+        "runs": [
+          262.3,
+          254.2,
+          254.5,
+          278.0,
+          260.2,
+          267.7,
+          253.4,
+          262.9,
+          259.5,
+          276.3,
+          252.2,
+          242.8,
+          280.9,
+          268.2,
+          253.6,
+          260.4,
+          264.4,
+          273.4,
+          259.1,
+          264.3,
+          258.9,
+          253.2,
+          263.8,
+          265.3,
+          263.8,
+          250.9,
+          256.8,
+          282.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/sustained/result.json b/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/sustained/result.json
index 19390809..a8b21777 100644
--- a/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/sustained/result.json
+++ b/results/verified/huawei_ascend_ascend910x16_suite_G_ascend_vllm_ascend_d4aa9fda_329a2b9e/sustained/result.json
@@ -522,7 +522,44 @@
       "sustained_throughput_tokens_per_sec": 262.2,
       "throttle_ratio": 0.861,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": -18.4
+      "ttft_p99_drift_ms": -18.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 262.2,
+        "std": 9.5,
+        "cv_pct": 3.62,
+        "stability": "noisy",
+        "runs": [
+          262.3,
+          254.2,
+          254.5,
+          278.0,
+          260.2,
+          267.7,
+          253.4,
+          262.9,
+          259.5,
+          276.3,
+          252.2,
+          242.8,
+          280.9,
+          268.2,
+          253.6,
+          260.4,
+          264.4,
+          273.4,
+          259.1,
+          264.3,
+          258.9,
+          253.2,
+          263.8,
+          265.3,
+          263.8,
+          250.9,
+          256.8,
+          282.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/result.json b/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/result.json
index 9d0b9088..aae2043c 100644
--- a/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/result.json
+++ b/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/result.json
@@ -620,7 +620,44 @@
       "sustained_throughput_tokens_per_sec": 376.6,
       "throttle_ratio": 0.893,
       "throttle_onset_minute": 21.0,
-      "ttft_p99_drift_ms": -8.0
+      "ttft_p99_drift_ms": -8.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 376.6,
+        "std": 11.4,
+        "cv_pct": 3.03,
+        "stability": "noisy",
+        "runs": [
+          368.1,
+          384.3,
+          387.7,
+          375.4,
+          366.8,
+          376.1,
+          382.0,
+          372.0,
+          377.9,
+          372.0,
+          369.6,
+          377.6,
+          376.8,
+          385.8,
+          362.3,
+          392.8,
+          363.8,
+          393.9,
+          373.8,
+          352.6,
+          394.1,
+          360.1,
+          390.1,
+          373.1,
+          388.4,
+          370.3,
+          394.7,
+          363.7
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/sustained/result.json b/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/sustained/result.json
index 0a93919d..f2d38380 100644
--- a/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/sustained/result.json
+++ b/results/verified/huawei_ascend_ascend910x1_suite_A_ascend_vllm_ascend_d4aa9fda_74d19743/sustained/result.json
@@ -522,7 +522,44 @@
       "sustained_throughput_tokens_per_sec": 376.6,
       "throttle_ratio": 0.893,
       "throttle_onset_minute": 21.0,
-      "ttft_p99_drift_ms": -8.0
+      "ttft_p99_drift_ms": -8.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 376.6,
+        "std": 11.4,
+        "cv_pct": 3.03,
+        "stability": "noisy",
+        "runs": [
+          368.1,
+          384.3,
+          387.7,
+          375.4,
+          366.8,
+          376.1,
+          382.0,
+          372.0,
+          377.9,
+          372.0,
+          369.6,
+          377.6,
+          376.8,
+          385.8,
+          362.3,
+          392.8,
+          363.8,
+          393.9,
+          373.8,
+          352.6,
+          394.1,
+          360.1,
+          390.1,
+          373.1,
+          388.4,
+          370.3,
+          394.7,
+          363.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/result.json b/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/result.json
index bc4cf352..7ee86f6b 100644
--- a/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/result.json
+++ b/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/result.json
@@ -564,7 +564,44 @@
       "sustained_throughput_tokens_per_sec": 54.2,
       "throttle_ratio": 0.784,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -321.6
+      "ttft_p99_drift_ms": -321.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 54.2,
+        "std": 3.8,
+        "cv_pct": 7.06,
+        "stability": "noisy",
+        "runs": [
+          46.9,
+          55.4,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.5,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.4,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.4,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.5,
+          55.5,
+          46.9,
+          59.8,
+          51.2,
+          59.8,
+          51.2
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/sustained/result.json b/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/sustained/result.json
index 03f09823..125e8f48 100644
--- a/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/sustained/result.json
+++ b/results/verified/huawei_ascend_ascend910x1_suite_D_ascend_vllm_ascend_d4aa9fda_6c1e7ffe/sustained/result.json
@@ -522,7 +522,44 @@
       "sustained_throughput_tokens_per_sec": 54.2,
       "throttle_ratio": 0.784,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -321.6
+      "ttft_p99_drift_ms": -321.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 54.2,
+        "std": 3.8,
+        "cv_pct": 7.06,
+        "stability": "noisy",
+        "runs": [
+          46.9,
+          55.4,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.5,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.4,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.4,
+          55.5,
+          51.2,
+          59.7,
+          51.2,
+          55.5,
+          55.5,
+          46.9,
+          59.8,
+          51.2,
+          59.8,
+          51.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/result.json b/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/result.json
index 6f66bb12..0d5e67ec 100644
--- a/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/result.json
+++ b/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/result.json
@@ -456,7 +456,30 @@
       "sustained_throughput_tokens_per_sec": 2217.9,
       "throttle_ratio": 0.94,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -217.5
+      "ttft_p99_drift_ms": -217.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2217.9,
+        "std": 34.3,
+        "cv_pct": 1.54,
+        "stability": "stable",
+        "runs": [
+          2121.0,
+          2228.7,
+          2201.3,
+          2198.4,
+          2215.7,
+          2232.6,
+          2225.4,
+          2220.6,
+          2187.0,
+          2241.5,
+          2245.4,
+          2250.2,
+          2255.4,
+          2227.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/sustained/result.json b/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/sustained/result.json
index 2d1b61e9..bcb9d743 100644
--- a/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/sustained/result.json
+++ b/results/verified/huawei_ascend_ascend910x1_suite_F_ascend_vllm_ascend_d4aa9fda_8826a63d/sustained/result.json
@@ -372,7 +372,30 @@
       "sustained_throughput_tokens_per_sec": 2217.9,
       "throttle_ratio": 0.94,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -217.5
+      "ttft_p99_drift_ms": -217.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2217.9,
+        "std": 34.3,
+        "cv_pct": 1.54,
+        "stability": "stable",
+        "runs": [
+          2121.0,
+          2228.7,
+          2201.3,
+          2198.4,
+          2215.7,
+          2232.6,
+          2225.4,
+          2220.6,
+          2187.0,
+          2241.5,
+          2245.4,
+          2250.2,
+          2255.4,
+          2227.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/result.json
index 653ebddc..70f42b4b 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/result.json
@@ -495,7 +495,44 @@
       "sustained_throughput_tokens_per_sec": 484.0,
       "throttle_ratio": 0.892,
       "throttle_onset_minute": 10.0,
-      "ttft_p99_drift_ms": -13.4
+      "ttft_p99_drift_ms": -13.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 484.0,
+        "std": 15.3,
+        "cv_pct": 3.16,
+        "stability": "noisy",
+        "runs": [
+          489.2,
+          480.6,
+          476.6,
+          474.6,
+          484.4,
+          490.8,
+          462.1,
+          485.0,
+          453.9,
+          462.0,
+          486.8,
+          454.3,
+          483.2,
+          467.2,
+          483.4,
+          480.4,
+          468.5,
+          501.4,
+          507.5,
+          482.9,
+          503.3,
+          498.0,
+          496.9,
+          494.1,
+          496.2,
+          479.2,
+          508.6,
+          501.7
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/sustained/result.json
index d553f1f1..7075969f 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_A_nvidia_vllm_47f5d58e_b4a92b30/sustained/result.json
@@ -397,7 +397,44 @@
       "sustained_throughput_tokens_per_sec": 484.0,
       "throttle_ratio": 0.892,
       "throttle_onset_minute": 10.0,
-      "ttft_p99_drift_ms": -13.4
+      "ttft_p99_drift_ms": -13.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 484.0,
+        "std": 15.3,
+        "cv_pct": 3.16,
+        "stability": "noisy",
+        "runs": [
+          489.2,
+          480.6,
+          476.6,
+          474.6,
+          484.4,
+          490.8,
+          462.1,
+          485.0,
+          453.9,
+          462.0,
+          486.8,
+          454.3,
+          483.2,
+          467.2,
+          483.4,
+          480.4,
+          468.5,
+          501.4,
+          507.5,
+          482.9,
+          503.3,
+          498.0,
+          496.9,
+          494.1,
+          496.2,
+          479.2,
+          508.6,
+          501.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/result.json
index fbc82d38..649a377c 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/result.json
@@ -356,7 +356,30 @@
       "sustained_throughput_tokens_per_sec": 491.9,
       "throttle_ratio": 0.898,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -144.4
+      "ttft_p99_drift_ms": -144.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 491.9,
+        "std": 14.0,
+        "cv_pct": 2.85,
+        "stability": "stable",
+        "runs": [
+          459.7,
+          507.7,
+          490.3,
+          491.6,
+          491.2,
+          504.1,
+          484.6,
+          487.8,
+          504.7,
+          481.5,
+          511.8,
+          475.1,
+          500.9,
+          496.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/sustained/result.json
index b9c8c6bf..1cab7952 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/bf16/sustained/result.json
@@ -247,7 +247,30 @@
       "sustained_throughput_tokens_per_sec": 491.9,
       "throttle_ratio": 0.898,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -144.4
+      "ttft_p99_drift_ms": -144.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 491.9,
+        "std": 14.0,
+        "cv_pct": 2.85,
+        "stability": "stable",
+        "runs": [
+          459.7,
+          507.7,
+          490.3,
+          491.6,
+          491.2,
+          504.1,
+          484.6,
+          487.8,
+          504.7,
+          481.5,
+          511.8,
+          475.1,
+          500.9,
+          496.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/result.json
index b98c6b18..7b184c01 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/result.json
@@ -356,7 +356,30 @@
       "sustained_throughput_tokens_per_sec": 709.1,
       "throttle_ratio": 0.935,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -196.6
+      "ttft_p99_drift_ms": -196.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 709.1,
+        "std": 12.2,
+        "cv_pct": 1.72,
+        "stability": "stable",
+        "runs": [
+          674.8,
+          719.2,
+          722.0,
+          704.1,
+          703.0,
+          722.0,
+          703.0,
+          708.7,
+          717.0,
+          717.1,
+          706.9,
+          716.7,
+          702.6,
+          710.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/sustained/result.json
index 554f0644..006c7372 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/fp8/sustained/result.json
@@ -247,7 +247,30 @@
       "sustained_throughput_tokens_per_sec": 709.1,
       "throttle_ratio": 0.935,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -196.6
+      "ttft_p99_drift_ms": -196.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 709.1,
+        "std": 12.2,
+        "cv_pct": 1.72,
+        "stability": "stable",
+        "runs": [
+          674.8,
+          719.2,
+          722.0,
+          704.1,
+          703.0,
+          722.0,
+          703.0,
+          708.7,
+          717.0,
+          717.1,
+          706.9,
+          716.7,
+          702.6,
+          710.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/result.json
index 67954308..20645b1e 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/result.json
@@ -356,7 +356,30 @@
       "sustained_throughput_tokens_per_sec": 813.5,
       "throttle_ratio": 0.926,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -204.3
+      "ttft_p99_drift_ms": -204.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 813.5,
+        "std": 17.7,
+        "cv_pct": 2.17,
+        "stability": "stable",
+        "runs": [
+          778.9,
+          820.5,
+          809.4,
+          823.7,
+          823.7,
+          793.7,
+          828.6,
+          841.5,
+          797.3,
+          818.5,
+          799.6,
+          829.7,
+          826.3,
+          797.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/sustained/result.json
index 36bb8403..69a88e9e 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w4a16/sustained/result.json
@@ -247,7 +247,30 @@
       "sustained_throughput_tokens_per_sec": 813.5,
       "throttle_ratio": 0.926,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -204.3
+      "ttft_p99_drift_ms": -204.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 813.5,
+        "std": 17.7,
+        "cv_pct": 2.17,
+        "stability": "stable",
+        "runs": [
+          778.9,
+          820.5,
+          809.4,
+          823.7,
+          823.7,
+          793.7,
+          828.6,
+          841.5,
+          797.3,
+          818.5,
+          799.6,
+          829.7,
+          826.3,
+          797.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/result.json
index ca1752d1..a9bc15bf 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/result.json
@@ -356,7 +356,30 @@
       "sustained_throughput_tokens_per_sec": 700.0,
       "throttle_ratio": 0.945,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -208.8
+      "ttft_p99_drift_ms": -208.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 700.0,
+        "std": 10.5,
+        "cv_pct": 1.51,
+        "stability": "stable",
+        "runs": [
+          676.5,
+          702.7,
+          701.1,
+          708.1,
+          693.3,
+          702.6,
+          708.2,
+          696.5,
+          697.5,
+          713.5,
+          685.0,
+          716.0,
+          702.4,
+          696.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/sustained/result.json
index 08e47dc3..e5d34321 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a16/sustained/result.json
@@ -247,7 +247,30 @@
       "sustained_throughput_tokens_per_sec": 700.0,
       "throttle_ratio": 0.945,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -208.8
+      "ttft_p99_drift_ms": -208.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 700.0,
+        "std": 10.5,
+        "cv_pct": 1.51,
+        "stability": "stable",
+        "runs": [
+          676.5,
+          702.7,
+          701.1,
+          708.1,
+          693.3,
+          702.6,
+          708.2,
+          696.5,
+          697.5,
+          713.5,
+          685.0,
+          716.0,
+          702.4,
+          696.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/result.json
index 040a4132..9f88f9ba 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/result.json
@@ -356,7 +356,30 @@
       "sustained_throughput_tokens_per_sec": 657.7,
       "throttle_ratio": 0.922,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -106.5
+      "ttft_p99_drift_ms": -106.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 657.7,
+        "std": 11.4,
+        "cv_pct": 1.73,
+        "stability": "stable",
+        "runs": [
+          632.3,
+          685.6,
+          656.9,
+          651.9,
+          656.6,
+          660.3,
+          662.7,
+          648.2,
+          660.2,
+          654.3,
+          661.2,
+          664.8,
+          654.3,
+          659.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/sustained/result.json
index 74ebc6a6..15e689f1 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_C_nvidia_vllm_47f5d58e_57cc3fdf/w8a8/sustained/result.json
@@ -247,7 +247,30 @@
       "sustained_throughput_tokens_per_sec": 657.7,
       "throttle_ratio": 0.922,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -106.5
+      "ttft_p99_drift_ms": -106.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 657.7,
+        "std": 11.4,
+        "cv_pct": 1.73,
+        "stability": "stable",
+        "runs": [
+          632.3,
+          685.6,
+          656.9,
+          651.9,
+          656.6,
+          660.3,
+          662.7,
+          648.2,
+          660.2,
+          654.3,
+          661.2,
+          664.8,
+          654.3,
+          659.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/result.json
index 0003f085..e444f214 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/result.json
@@ -439,7 +439,44 @@
       "sustained_throughput_tokens_per_sec": 57.0,
       "throttle_ratio": 0.705,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -451.2
+      "ttft_p99_drift_ms": -451.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 57.0,
+        "std": 7.4,
+        "cv_pct": 13.07,
+        "stability": "high-variance",
+        "runs": [
+          51.2,
+          55.5,
+          68.2,
+          51.2,
+          51.2,
+          55.4,
+          72.5,
+          51.2,
+          55.5,
+          51.2,
+          55.5,
+          68.2,
+          55.5,
+          55.5,
+          51.1,
+          68.4,
+          55.4,
+          55.5,
+          51.2,
+          51.2,
+          72.5,
+          55.5,
+          51.2,
+          55.5,
+          51.2,
+          72.5,
+          51.2,
+          55.5
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/sustained/result.json
index d4e23453..75aad313 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_D_nvidia_vllm_47f5d58e_8e114cbe/sustained/result.json
@@ -397,7 +397,44 @@
       "sustained_throughput_tokens_per_sec": 57.0,
       "throttle_ratio": 0.705,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -451.2
+      "ttft_p99_drift_ms": -451.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 57.0,
+        "std": 7.4,
+        "cv_pct": 13.07,
+        "stability": "high-variance",
+        "runs": [
+          51.2,
+          55.5,
+          68.2,
+          51.2,
+          51.2,
+          55.4,
+          72.5,
+          51.2,
+          55.5,
+          51.2,
+          55.5,
+          68.2,
+          55.5,
+          55.5,
+          51.1,
+          68.4,
+          55.4,
+          55.5,
+          51.2,
+          51.2,
+          72.5,
+          55.5,
+          51.2,
+          55.5,
+          51.2,
+          72.5,
+          51.2,
+          55.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/result.json
index 23b30f8b..5e8ac339 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/result.json
@@ -331,7 +331,30 @@
       "sustained_throughput_tokens_per_sec": 3972.5,
       "throttle_ratio": 0.963,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -114.8
+      "ttft_p99_drift_ms": -114.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 3972.5,
+        "std": 41.8,
+        "cv_pct": 1.05,
+        "stability": "stable",
+        "runs": [
+          3951.2,
+          3922.5,
+          4011.2,
+          4072.5,
+          3992.6,
+          3964.1,
+          4012.5,
+          3989.7,
+          3938.4,
+          3932.9,
+          3939.0,
+          3938.5,
+          3995.8,
+          3954.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/sustained/result.json
index 445a5911..be0a3b44 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx1_suite_F_nvidia_vllm_47f5d58e_fe3156b5/sustained/result.json
@@ -247,7 +247,30 @@
       "sustained_throughput_tokens_per_sec": 3972.5,
       "throttle_ratio": 0.963,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -114.8
+      "ttft_p99_drift_ms": -114.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 3972.5,
+        "std": 41.8,
+        "cv_pct": 1.05,
+        "stability": "stable",
+        "runs": [
+          3951.2,
+          3922.5,
+          4011.2,
+          4072.5,
+          3992.6,
+          3964.1,
+          4012.5,
+          3989.7,
+          3938.4,
+          3932.9,
+          3939.0,
+          3938.5,
+          3995.8,
+          3954.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/result.json b/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/result.json
index 7b24e63d..65a2c792 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/result.json
@@ -574,7 +574,44 @@
       "sustained_throughput_tokens_per_sec": 164.3,
       "throttle_ratio": 0.806,
       "throttle_onset_minute": 6.0,
-      "ttft_p99_drift_ms": 17.4
+      "ttft_p99_drift_ms": 17.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 164.3,
+        "std": 8.5,
+        "cv_pct": 5.2,
+        "stability": "noisy",
+        "runs": [
+          167.7,
+          172.9,
+          181.0,
+          164.4,
+          162.7,
+          175.0,
+          174.0,
+          173.0,
+          174.6,
+          164.4,
+          171.3,
+          166.6,
+          157.8,
+          160.2,
+          155.9,
+          166.3,
+          149.8,
+          161.3,
+          157.5,
+          171.4,
+          150.8,
+          162.6,
+          157.2,
+          160.2,
+          156.9,
+          168.6,
+          145.8,
+          170.8
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 81.65,
diff --git a/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/sustained/result.json
index 94afb943..33933391 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx8_suite_B_nvidia_vllm_47f5d58e_14410aea/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 164.3,
       "throttle_ratio": 0.806,
       "throttle_onset_minute": 6.0,
-      "ttft_p99_drift_ms": 17.4
+      "ttft_p99_drift_ms": 17.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 164.3,
+        "std": 8.5,
+        "cv_pct": 5.2,
+        "stability": "noisy",
+        "runs": [
+          167.7,
+          172.9,
+          181.0,
+          164.4,
+          162.7,
+          175.0,
+          174.0,
+          173.0,
+          174.6,
+          164.4,
+          171.3,
+          166.6,
+          157.8,
+          160.2,
+          155.9,
+          166.3,
+          149.8,
+          161.3,
+          157.5,
+          171.4,
+          150.8,
+          162.6,
+          157.2,
+          160.2,
+          156.9,
+          168.6,
+          145.8,
+          170.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/result.json b/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/result.json
index 993f6d90..fe47c5a0 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/result.json
@@ -571,7 +571,44 @@
       "sustained_throughput_tokens_per_sec": 472.7,
       "throttle_ratio": 0.902,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -2.3
+      "ttft_p99_drift_ms": -2.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 472.7,
+        "std": 11.2,
+        "cv_pct": 2.38,
+        "stability": "stable",
+        "runs": [
+          482.4,
+          470.0,
+          478.0,
+          464.6,
+          475.5,
+          474.1,
+          475.7,
+          467.2,
+          478.8,
+          476.8,
+          443.2,
+          477.3,
+          480.2,
+          459.0,
+          465.2,
+          486.8,
+          468.0,
+          466.5,
+          484.3,
+          454.4,
+          491.3,
+          483.4,
+          482.3,
+          454.0,
+          481.7,
+          470.6,
+          462.2,
+          482.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/sustained/result.json b/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/sustained/result.json
index ddf5c312..c6564030 100644
--- a/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_40gbx8_suite_G_nvidia_vllm_47f5d58e_08de2dc2/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 472.7,
       "throttle_ratio": 0.902,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -2.3
+      "ttft_p99_drift_ms": -2.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 472.7,
+        "std": 11.2,
+        "cv_pct": 2.38,
+        "stability": "stable",
+        "runs": [
+          482.4,
+          470.0,
+          478.0,
+          464.6,
+          475.5,
+          474.1,
+          475.7,
+          467.2,
+          478.8,
+          476.8,
+          443.2,
+          477.3,
+          480.2,
+          459.0,
+          465.2,
+          486.8,
+          468.0,
+          466.5,
+          484.3,
+          454.4,
+          491.3,
+          483.4,
+          482.3,
+          454.0,
+          481.7,
+          470.6,
+          462.2,
+          482.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/result.json
index ca26d93b..1dec8d50 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/result.json
@@ -494,7 +494,44 @@
       "sustained_throughput_tokens_per_sec": 712.3,
       "throttle_ratio": 0.947,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -0.2
+      "ttft_p99_drift_ms": -0.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 712.3,
+        "std": 9.4,
+        "cv_pct": 1.32,
+        "stability": "stable",
+        "runs": [
+          707.4,
+          710.2,
+          728.3,
+          707.4,
+          711.2,
+          708.9,
+          701.6,
+          723.0,
+          702.0,
+          716.4,
+          718.5,
+          716.5,
+          708.7,
+          711.1,
+          715.6,
+          699.2,
+          721.0,
+          689.9,
+          718.4,
+          720.0,
+          726.1,
+          713.8,
+          694.3,
+          709.4,
+          720.1,
+          714.8,
+          705.6,
+          725.1
+        ]
+      }
     },
     "burst": {
       "sla_ttft_ms": 500,
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/sustained/result.json
index 1fc95fb6..c99db7f5 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm020_0f6c56e4_8f83bfab/sustained/result.json
@@ -394,7 +394,44 @@
       "sustained_throughput_tokens_per_sec": 712.3,
       "throttle_ratio": 0.947,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -0.2
+      "ttft_p99_drift_ms": -0.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 712.3,
+        "std": 9.4,
+        "cv_pct": 1.32,
+        "stability": "stable",
+        "runs": [
+          707.4,
+          710.2,
+          728.3,
+          707.4,
+          711.2,
+          708.9,
+          701.6,
+          723.0,
+          702.0,
+          716.4,
+          718.5,
+          716.5,
+          708.7,
+          711.1,
+          715.6,
+          699.2,
+          721.0,
+          689.9,
+          718.4,
+          720.0,
+          726.1,
+          713.8,
+          694.3,
+          709.4,
+          720.1,
+          714.8,
+          705.6,
+          725.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/result.json
index 10757712..f071d27d 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/result.json
@@ -500,7 +500,44 @@
       "sustained_throughput_tokens_per_sec": 551.9,
       "throttle_ratio": 0.868,
       "throttle_onset_minute": 18.0,
-      "ttft_p99_drift_ms": 9.0
+      "ttft_p99_drift_ms": 9.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 551.9,
+        "std": 25.7,
+        "cv_pct": 4.66,
+        "stability": "noisy",
+        "runs": [
+          566.9,
+          585.4,
+          575.4,
+          575.6,
+          584.1,
+          571.1,
+          568.6,
+          582.2,
+          558.0,
+          562.9,
+          581.4,
+          575.8,
+          578.3,
+          572.0,
+          568.3,
+          542.1,
+          526.1,
+          528.9,
+          541.5,
+          521.5,
+          513.8,
+          536.7,
+          510.0,
+          536.8,
+          527.1,
+          508.2,
+          521.6,
+          531.8
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/sustained/result.json
index c1e56599..46ed843d 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_ed4b0557/sustained/result.json
@@ -402,7 +402,44 @@
       "sustained_throughput_tokens_per_sec": 551.9,
       "throttle_ratio": 0.868,
       "throttle_onset_minute": 18.0,
-      "ttft_p99_drift_ms": 9.0
+      "ttft_p99_drift_ms": 9.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 551.9,
+        "std": 25.7,
+        "cv_pct": 4.66,
+        "stability": "noisy",
+        "runs": [
+          566.9,
+          585.4,
+          575.4,
+          575.6,
+          584.1,
+          571.1,
+          568.6,
+          582.2,
+          558.0,
+          562.9,
+          581.4,
+          575.8,
+          578.3,
+          572.0,
+          568.3,
+          542.1,
+          526.1,
+          528.9,
+          541.5,
+          521.5,
+          513.8,
+          536.7,
+          510.0,
+          536.8,
+          527.1,
+          508.2,
+          521.6,
+          531.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/result.json
index 75e68ff6..eefbf857 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/result.json
@@ -357,7 +357,30 @@
       "sustained_throughput_tokens_per_sec": 706.9,
       "throttle_ratio": 0.899,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -360.1
+      "ttft_p99_drift_ms": -360.1,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 706.9,
+        "std": 19.3,
+        "cv_pct": 2.73,
+        "stability": "stable",
+        "runs": [
+          655.0,
+          711.0,
+          698.3,
+          718.8,
+          724.4,
+          701.8,
+          706.9,
+          720.6,
+          697.1,
+          726.2,
+          701.9,
+          728.3,
+          688.9,
+          717.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/sustained/result.json
index ee4ebabf..1a000d46 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/bf16/sustained/result.json
@@ -244,7 +244,30 @@
       "sustained_throughput_tokens_per_sec": 706.9,
       "throttle_ratio": 0.899,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -360.1
+      "ttft_p99_drift_ms": -360.1,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 706.9,
+        "std": 19.3,
+        "cv_pct": 2.73,
+        "stability": "stable",
+        "runs": [
+          655.0,
+          711.0,
+          698.3,
+          718.8,
+          724.4,
+          701.8,
+          706.9,
+          720.6,
+          697.1,
+          726.2,
+          701.9,
+          728.3,
+          688.9,
+          717.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/result.json
index 27e0744a..65f49b8f 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/result.json
@@ -362,7 +362,30 @@
       "sustained_throughput_tokens_per_sec": 437.3,
       "throttle_ratio": 0.897,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -632.2
+      "ttft_p99_drift_ms": -632.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 437.3,
+        "std": 11.6,
+        "cv_pct": 2.66,
+        "stability": "stable",
+        "runs": [
+          409.4,
+          431.5,
+          456.2,
+          439.7,
+          430.2,
+          432.7,
+          452.2,
+          436.1,
+          432.3,
+          431.9,
+          449.0,
+          445.3,
+          441.9,
+          433.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/sustained/result.json
index 9982bb06..ab83ad8d 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w4a16/sustained/result.json
@@ -249,7 +249,30 @@
       "sustained_throughput_tokens_per_sec": 437.3,
       "throttle_ratio": 0.897,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -632.2
+      "ttft_p99_drift_ms": -632.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 437.3,
+        "std": 11.6,
+        "cv_pct": 2.66,
+        "stability": "stable",
+        "runs": [
+          409.4,
+          431.5,
+          456.2,
+          439.7,
+          430.2,
+          432.7,
+          452.2,
+          436.1,
+          432.3,
+          431.9,
+          449.0,
+          445.3,
+          441.9,
+          433.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/result.json
index 485a0fb3..feea95d5 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/result.json
@@ -362,7 +362,30 @@
       "sustained_throughput_tokens_per_sec": 494.1,
       "throttle_ratio": 0.905,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -320.5
+      "ttft_p99_drift_ms": -320.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 494.1,
+        "std": 12.0,
+        "cv_pct": 2.42,
+        "stability": "stable",
+        "runs": [
+          456.8,
+          504.0,
+          486.6,
+          499.0,
+          496.0,
+          498.3,
+          495.4,
+          496.1,
+          503.2,
+          489.6,
+          504.8,
+          500.4,
+          494.6,
+          492.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/sustained/result.json
index 0fa36d8f..f1ef7eaa 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a16/sustained/result.json
@@ -249,7 +249,30 @@
       "sustained_throughput_tokens_per_sec": 494.1,
       "throttle_ratio": 0.905,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -320.5
+      "ttft_p99_drift_ms": -320.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 494.1,
+        "std": 12.0,
+        "cv_pct": 2.42,
+        "stability": "stable",
+        "runs": [
+          456.8,
+          504.0,
+          486.6,
+          499.0,
+          496.0,
+          498.3,
+          495.4,
+          496.1,
+          503.2,
+          489.6,
+          504.8,
+          500.4,
+          494.6,
+          492.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/result.json
index ff6e59c1..0828f267 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/result.json
@@ -362,7 +362,30 @@
       "sustained_throughput_tokens_per_sec": 399.4,
       "throttle_ratio": 0.879,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -331.5
+      "ttft_p99_drift_ms": -331.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 399.4,
+        "std": 12.5,
+        "cv_pct": 3.13,
+        "stability": "noisy",
+        "runs": [
+          366.9,
+          402.7,
+          400.8,
+          402.1,
+          395.3,
+          398.4,
+          410.0,
+          398.6,
+          390.3,
+          417.2,
+          391.1,
+          408.7,
+          415.2,
+          395.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/sustained/result.json
index eaea3a87..01b72946 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm020_0f6c56e4_ffd81462/w8a8/sustained/result.json
@@ -249,7 +249,30 @@
       "sustained_throughput_tokens_per_sec": 399.4,
       "throttle_ratio": 0.879,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -331.5
+      "ttft_p99_drift_ms": -331.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 399.4,
+        "std": 12.5,
+        "cv_pct": 3.13,
+        "stability": "noisy",
+        "runs": [
+          366.9,
+          402.7,
+          400.8,
+          402.1,
+          395.3,
+          398.4,
+          410.0,
+          398.6,
+          390.3,
+          417.2,
+          391.1,
+          408.7,
+          415.2,
+          395.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/result.json
index f1aeef5e..1db3621f 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 537.9,
       "throttle_ratio": 0.87,
       "throttle_onset_minute": 9.0,
-      "ttft_p99_drift_ms": -138.7
+      "ttft_p99_drift_ms": -138.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 537.9,
+        "std": 25.5,
+        "cv_pct": 4.74,
+        "stability": "noisy",
+        "runs": [
+          516.9,
+          570.2,
+          563.2,
+          574.0,
+          567.1,
+          569.5,
+          533.9,
+          526.7,
+          509.9,
+          532.6,
+          499.2,
+          522.4,
+          520.7,
+          524.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/sustained/result.json
index 74ec1b43..3948212c 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/bf16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 537.9,
       "throttle_ratio": 0.87,
       "throttle_onset_minute": 9.0,
-      "ttft_p99_drift_ms": -138.7
+      "ttft_p99_drift_ms": -138.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 537.9,
+        "std": 25.5,
+        "cv_pct": 4.74,
+        "stability": "noisy",
+        "runs": [
+          516.9,
+          570.2,
+          563.2,
+          574.0,
+          567.1,
+          569.5,
+          533.9,
+          526.7,
+          509.9,
+          532.6,
+          499.2,
+          522.4,
+          520.7,
+          524.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/result.json
index 5d3e2c1d..592ad309 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 709.7,
       "throttle_ratio": 0.927,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -189.7
+      "ttft_p99_drift_ms": -189.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 709.7,
+        "std": 15.4,
+        "cv_pct": 2.18,
+        "stability": "stable",
+        "runs": [
+          741.9,
+          718.6,
+          701.6,
+          688.1,
+          729.3,
+          691.6,
+          728.3,
+          715.4,
+          712.9,
+          702.4,
+          699.8,
+          698.3,
+          705.5,
+          702.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/sustained/result.json
index c8f36da6..0a71eb6e 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/fp8/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 709.7,
       "throttle_ratio": 0.927,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -189.7
+      "ttft_p99_drift_ms": -189.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 709.7,
+        "std": 15.4,
+        "cv_pct": 2.18,
+        "stability": "stable",
+        "runs": [
+          741.9,
+          718.6,
+          701.6,
+          688.1,
+          729.3,
+          691.6,
+          728.3,
+          715.4,
+          712.9,
+          702.4,
+          699.8,
+          698.3,
+          705.5,
+          702.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/result.json
index d4c5752a..fc10a331 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 757.0,
       "throttle_ratio": 0.942,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -185.6
+      "ttft_p99_drift_ms": -185.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 757.0,
+        "std": 14.6,
+        "cv_pct": 1.92,
+        "stability": "stable",
+        "runs": [
+          737.0,
+          764.7,
+          762.2,
+          771.5,
+          740.1,
+          755.5,
+          752.4,
+          744.3,
+          781.3,
+          745.1,
+          782.2,
+          748.8,
+          747.8,
+          765.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/sustained/result.json
index 46c58f19..afa0d716 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w4a16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 757.0,
       "throttle_ratio": 0.942,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -185.6
+      "ttft_p99_drift_ms": -185.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 757.0,
+        "std": 14.6,
+        "cv_pct": 1.92,
+        "stability": "stable",
+        "runs": [
+          737.0,
+          764.7,
+          762.2,
+          771.5,
+          740.1,
+          755.5,
+          752.4,
+          744.3,
+          781.3,
+          745.1,
+          782.2,
+          748.8,
+          747.8,
+          765.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/result.json
index d2f40e9b..de77ce59 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 694.8,
       "throttle_ratio": 0.912,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -179.3
+      "ttft_p99_drift_ms": -179.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 694.8,
+        "std": 17.4,
+        "cv_pct": 2.51,
+        "stability": "stable",
+        "runs": [
+          728.5,
+          716.1,
+          677.7,
+          692.5,
+          712.4,
+          664.6,
+          700.4,
+          688.4,
+          678.1,
+          692.0,
+          710.5,
+          679.9,
+          690.7,
+          695.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/sustained/result.json
index 21592a14..1710fbb0 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 694.8,
       "throttle_ratio": 0.912,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -179.3
+      "ttft_p99_drift_ms": -179.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 694.8,
+        "std": 17.4,
+        "cv_pct": 2.51,
+        "stability": "stable",
+        "runs": [
+          728.5,
+          716.1,
+          677.7,
+          692.5,
+          712.4,
+          664.6,
+          700.4,
+          688.4,
+          678.1,
+          692.0,
+          710.5,
+          679.9,
+          690.7,
+          695.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/result.json
index 0daadb7b..4e391997 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 643.9,
       "throttle_ratio": 0.925,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -92.0
+      "ttft_p99_drift_ms": -92.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 643.9,
+        "std": 13.7,
+        "cv_pct": 2.12,
+        "stability": "stable",
+        "runs": [
+          655.9,
+          662.5,
+          620.7,
+          649.8,
+          643.7,
+          631.6,
+          645.2,
+          628.6,
+          671.1,
+          634.4,
+          650.0,
+          633.5,
+          643.4,
+          643.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/sustained/result.json
index bf2083e9..5685fadc 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_6940965a/w8a8/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 643.9,
       "throttle_ratio": 0.925,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -92.0
+      "ttft_p99_drift_ms": -92.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 643.9,
+        "std": 13.7,
+        "cv_pct": 2.12,
+        "stability": "stable",
+        "runs": [
+          655.9,
+          662.5,
+          620.7,
+          649.8,
+          643.7,
+          631.6,
+          645.2,
+          628.6,
+          671.1,
+          634.4,
+          650.0,
+          633.5,
+          643.4,
+          643.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/result.json
index ac745b18..fa0ebd2e 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/result.json
@@ -438,7 +438,44 @@
       "sustained_throughput_tokens_per_sec": 58.7,
       "throttle_ratio": 0.866,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 202.0
+      "ttft_p99_drift_ms": 202.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 58.7,
+        "std": 3.2,
+        "cv_pct": 5.46,
+        "stability": "noisy",
+        "runs": [
+          55.4,
+          59.7,
+          55.5,
+          64.0,
+          55.5,
+          59.7,
+          59.8,
+          59.7,
+          59.7,
+          55.5,
+          64.0,
+          55.5,
+          55.5,
+          59.7,
+          64.0,
+          55.4,
+          55.5,
+          64.0,
+          55.5,
+          59.7,
+          55.5,
+          64.0,
+          55.4,
+          59.8,
+          59.7,
+          59.7,
+          59.7,
+          55.5
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/sustained/result.json
index 097e0e91..23ccbdfa 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm020_0f6c56e4_43e96189/sustained/result.json
@@ -394,7 +394,44 @@
       "sustained_throughput_tokens_per_sec": 58.7,
       "throttle_ratio": 0.866,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 202.0
+      "ttft_p99_drift_ms": 202.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 58.7,
+        "std": 3.2,
+        "cv_pct": 5.46,
+        "stability": "noisy",
+        "runs": [
+          55.4,
+          59.7,
+          55.5,
+          64.0,
+          55.5,
+          59.7,
+          59.8,
+          59.7,
+          59.7,
+          55.5,
+          64.0,
+          55.5,
+          55.5,
+          59.7,
+          64.0,
+          55.4,
+          55.5,
+          64.0,
+          55.5,
+          59.7,
+          55.5,
+          64.0,
+          55.4,
+          59.8,
+          59.7,
+          59.7,
+          59.7,
+          55.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/result.json
index 57b288a8..b72baa3a 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/result.json
@@ -444,7 +444,44 @@
       "sustained_throughput_tokens_per_sec": 67.1,
       "throttle_ratio": 0.501,
       "throttle_onset_minute": 21.0,
-      "ttft_p99_drift_ms": -474.5
+      "ttft_p99_drift_ms": -474.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 67.1,
+        "std": 6.4,
+        "cv_pct": 9.6,
+        "stability": "high-variance",
+        "runs": [
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          34.2,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.2
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/sustained/result.json
index bf336ef5..aa42e23f 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_7bef8eef/sustained/result.json
@@ -402,7 +402,44 @@
       "sustained_throughput_tokens_per_sec": 67.1,
       "throttle_ratio": 0.501,
       "throttle_onset_minute": 21.0,
-      "ttft_p99_drift_ms": -474.5
+      "ttft_p99_drift_ms": -474.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 67.1,
+        "std": 6.4,
+        "cv_pct": 9.6,
+        "stability": "high-variance",
+        "runs": [
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          34.2,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/result.json
index 2e7e0ce3..dc61504c 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/result.json
@@ -336,7 +336,30 @@
       "sustained_throughput_tokens_per_sec": 11576.2,
       "throttle_ratio": 0.958,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -15.4
+      "ttft_p99_drift_ms": -15.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 11576.2,
+        "std": 139.5,
+        "cv_pct": 1.2,
+        "stability": "stable",
+        "runs": [
+          11541.5,
+          11672.6,
+          11721.9,
+          11526.8,
+          11228.1,
+          11380.9,
+          11711.7,
+          11643.5,
+          11662.7,
+          11612.6,
+          11623.8,
+          11639.3,
+          11651.4,
+          11450.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/sustained/result.json
index 6851ff63..78b09b49 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm020_0f6c56e4_a4e6a6e4/sustained/result.json
@@ -249,7 +249,30 @@
       "sustained_throughput_tokens_per_sec": 11576.2,
       "throttle_ratio": 0.958,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -15.4
+      "ttft_p99_drift_ms": -15.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 11576.2,
+        "std": 139.5,
+        "cv_pct": 1.2,
+        "stability": "stable",
+        "runs": [
+          11541.5,
+          11672.6,
+          11721.9,
+          11526.8,
+          11228.1,
+          11380.9,
+          11711.7,
+          11643.5,
+          11662.7,
+          11612.6,
+          11623.8,
+          11639.3,
+          11651.4,
+          11450.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/result.json
index 4d2726b7..c9818acf 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/result.json
@@ -336,7 +336,30 @@
       "sustained_throughput_tokens_per_sec": 2386.8,
       "throttle_ratio": 0.746,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -9.6
+      "ttft_p99_drift_ms": -9.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2386.8,
+        "std": 225.6,
+        "cv_pct": 9.45,
+        "stability": "high-variance",
+        "runs": [
+          2796.4,
+          2979.6,
+          2296.3,
+          2250.9,
+          2332.1,
+          2333.1,
+          2362.3,
+          2258.5,
+          2439.4,
+          2412.5,
+          2221.4,
+          2241.1,
+          2268.2,
+          2223.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/sustained/result.json b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/sustained/result.json
index a1bb052b..27d853c5 100644
--- a/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/sustained/result.json
+++ b/results/verified/nvidia_a100_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_52ad2fe3/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 2386.8,
       "throttle_ratio": 0.746,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -9.6
+      "ttft_p99_drift_ms": -9.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2386.8,
+        "std": 225.6,
+        "cv_pct": 9.45,
+        "stability": "high-variance",
+        "runs": [
+          2796.4,
+          2979.6,
+          2296.3,
+          2250.9,
+          2332.1,
+          2333.1,
+          2362.3,
+          2258.5,
+          2439.4,
+          2412.5,
+          2221.4,
+          2241.1,
+          2268.2,
+          2223.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/result.json
index 09fc23d5..56b152e2 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/result.json
@@ -500,7 +500,44 @@
       "sustained_throughput_tokens_per_sec": 546.4,
       "throttle_ratio": 0.881,
       "throttle_onset_minute": 23.0,
-      "ttft_p99_drift_ms": -0.7
+      "ttft_p99_drift_ms": -0.7,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 546.4,
+        "std": 17.4,
+        "cv_pct": 3.19,
+        "stability": "noisy",
+        "runs": [
+          566.8,
+          559.7,
+          544.8,
+          551.0,
+          525.8,
+          529.8,
+          531.9,
+          554.8,
+          584.0,
+          548.1,
+          572.2,
+          531.0,
+          546.0,
+          552.1,
+          545.3,
+          565.6,
+          535.0,
+          550.0,
+          552.7,
+          555.4,
+          540.9,
+          518.7,
+          514.7,
+          542.3,
+          546.4,
+          521.4,
+          578.2,
+          533.5
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/sustained/result.json
index d327a16a..f3bb99cb 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_A_nvidia_vllm_47f5d58e_298e6500/sustained/result.json
@@ -402,7 +402,44 @@
       "sustained_throughput_tokens_per_sec": 546.4,
       "throttle_ratio": 0.881,
       "throttle_onset_minute": 23.0,
-      "ttft_p99_drift_ms": -0.7
+      "ttft_p99_drift_ms": -0.7,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 546.4,
+        "std": 17.4,
+        "cv_pct": 3.19,
+        "stability": "noisy",
+        "runs": [
+          566.8,
+          559.7,
+          544.8,
+          551.0,
+          525.8,
+          529.8,
+          531.9,
+          554.8,
+          584.0,
+          548.1,
+          572.2,
+          531.0,
+          546.0,
+          552.1,
+          545.3,
+          565.6,
+          535.0,
+          550.0,
+          552.7,
+          555.4,
+          540.9,
+          518.7,
+          514.7,
+          542.3,
+          546.4,
+          521.4,
+          578.2,
+          533.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/result.json
index 0ed01fcb..8d42221b 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 534.5,
       "throttle_ratio": 0.933,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -140.3
+      "ttft_p99_drift_ms": -140.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 534.5,
+        "std": 12.1,
+        "cv_pct": 2.26,
+        "stability": "stable",
+        "runs": [
+          559.4,
+          525.8,
+          552.5,
+          528.7,
+          547.5,
+          525.4,
+          532.6,
+          526.2,
+          521.7,
+          531.5,
+          521.7,
+          533.9,
+          528.7,
+          546.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/sustained/result.json
index 7b8752c6..36af61f6 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/bf16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 534.5,
       "throttle_ratio": 0.933,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -140.3
+      "ttft_p99_drift_ms": -140.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 534.5,
+        "std": 12.1,
+        "cv_pct": 2.26,
+        "stability": "stable",
+        "runs": [
+          559.4,
+          525.8,
+          552.5,
+          528.7,
+          547.5,
+          525.4,
+          532.6,
+          526.2,
+          521.7,
+          531.5,
+          521.7,
+          533.9,
+          528.7,
+          546.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/result.json
index 73b517f2..63d529d9 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 760.1,
       "throttle_ratio": 0.893,
       "throttle_onset_minute": 10.0,
-      "ttft_p99_drift_ms": -171.7
+      "ttft_p99_drift_ms": -171.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 760.1,
+        "std": 33.5,
+        "cv_pct": 4.4,
+        "stability": "noisy",
+        "runs": [
+          769.6,
+          793.5,
+          770.5,
+          800.0,
+          794.2,
+          803.6,
+          796.6,
+          754.2,
+          723.9,
+          717.3,
+          739.1,
+          737.1,
+          719.0,
+          722.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/sustained/result.json
index 9f6d234c..bbe68d18 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/fp8/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 760.1,
       "throttle_ratio": 0.893,
       "throttle_onset_minute": 10.0,
-      "ttft_p99_drift_ms": -171.7
+      "ttft_p99_drift_ms": -171.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 760.1,
+        "std": 33.5,
+        "cv_pct": 4.4,
+        "stability": "noisy",
+        "runs": [
+          769.6,
+          793.5,
+          770.5,
+          800.0,
+          794.2,
+          803.6,
+          796.6,
+          754.2,
+          723.9,
+          717.3,
+          739.1,
+          737.1,
+          719.0,
+          722.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/result.json
index 3a45077f..68f0f389 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 829.8,
       "throttle_ratio": 0.83,
       "throttle_onset_minute": 7.0,
-      "ttft_p99_drift_ms": -193.6
+      "ttft_p99_drift_ms": -193.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 829.8,
+        "std": 56.4,
+        "cv_pct": 6.8,
+        "stability": "noisy",
+        "runs": [
+          836.7,
+          902.0,
+          882.7,
+          918.5,
+          861.0,
+          915.1,
+          763.1,
+          803.2,
+          805.7,
+          783.7,
+          818.0,
+          791.1,
+          772.9,
+          762.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/sustained/result.json
index 587e4d7e..0bd06286 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w4a16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 829.8,
       "throttle_ratio": 0.83,
       "throttle_onset_minute": 7.0,
-      "ttft_p99_drift_ms": -193.6
+      "ttft_p99_drift_ms": -193.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 829.8,
+        "std": 56.4,
+        "cv_pct": 6.8,
+        "stability": "noisy",
+        "runs": [
+          836.7,
+          902.0,
+          882.7,
+          918.5,
+          861.0,
+          915.1,
+          763.1,
+          803.2,
+          805.7,
+          783.7,
+          818.0,
+          791.1,
+          772.9,
+          762.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/result.json
index e207171a..ac12dd14 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 745.0,
       "throttle_ratio": 0.866,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": -190.8
+      "ttft_p99_drift_ms": -190.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 745.0,
+        "std": 40.9,
+        "cv_pct": 5.49,
+        "stability": "noisy",
+        "runs": [
+          738.9,
+          809.0,
+          793.6,
+          810.7,
+          771.9,
+          783.8,
+          735.7,
+          702.0,
+          702.2,
+          717.8,
+          702.5,
+          722.4,
+          702.6,
+          737.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/sustained/result.json
index bc45102a..ff160bbc 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 745.0,
       "throttle_ratio": 0.866,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": -190.8
+      "ttft_p99_drift_ms": -190.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 745.0,
+        "std": 40.9,
+        "cv_pct": 5.49,
+        "stability": "noisy",
+        "runs": [
+          738.9,
+          809.0,
+          793.6,
+          810.7,
+          771.9,
+          783.8,
+          735.7,
+          702.0,
+          702.2,
+          717.8,
+          702.5,
+          722.4,
+          702.6,
+          737.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/result.json
index 494d6c51..57802246 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 715.7,
       "throttle_ratio": 0.843,
       "throttle_onset_minute": 10.0,
-      "ttft_p99_drift_ms": -94.9
+      "ttft_p99_drift_ms": -94.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 715.7,
+        "std": 43.7,
+        "cv_pct": 6.1,
+        "stability": "noisy",
+        "runs": [
+          717.7,
+          704.7,
+          752.5,
+          762.7,
+          771.9,
+          752.2,
+          759.1,
+          732.5,
+          743.4,
+          681.3,
+          653.7,
+          669.9,
+          667.5,
+          650.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/sustained/result.json
index f37b0c8d..9ae0eda1 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_C_nvidia_vllm_47f5d58e_944773aa/w8a8/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 715.7,
       "throttle_ratio": 0.843,
       "throttle_onset_minute": 10.0,
-      "ttft_p99_drift_ms": -94.9
+      "ttft_p99_drift_ms": -94.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 715.7,
+        "std": 43.7,
+        "cv_pct": 6.1,
+        "stability": "noisy",
+        "runs": [
+          717.7,
+          704.7,
+          752.5,
+          762.7,
+          771.9,
+          752.2,
+          759.1,
+          732.5,
+          743.4,
+          681.3,
+          653.7,
+          669.9,
+          667.5,
+          650.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/result.json
index 5596dda6..07871c94 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/result.json
@@ -444,7 +444,44 @@
       "sustained_throughput_tokens_per_sec": 67.0,
       "throttle_ratio": 0.502,
       "throttle_onset_minute": 28.0,
-      "ttft_p99_drift_ms": -397.5
+      "ttft_p99_drift_ms": -397.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 67.0,
+        "std": 6.4,
+        "cv_pct": 9.57,
+        "stability": "high-variance",
+        "runs": [
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          67.9,
+          34.3,
+          68.3
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/sustained/result.json
index 108f892d..bbb4b142 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_D_nvidia_vllm_47f5d58e_4d0e7990/sustained/result.json
@@ -402,7 +402,44 @@
       "sustained_throughput_tokens_per_sec": 67.0,
       "throttle_ratio": 0.502,
       "throttle_onset_minute": 28.0,
-      "ttft_p99_drift_ms": -397.5
+      "ttft_p99_drift_ms": -397.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 67.0,
+        "std": 6.4,
+        "cv_pct": 9.57,
+        "stability": "high-variance",
+        "runs": [
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.3,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          68.2,
+          68.3,
+          67.9,
+          34.3,
+          68.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/result.json
index cc527f3c..d52a285a 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/result.json
@@ -336,7 +336,30 @@
       "sustained_throughput_tokens_per_sec": 2804.8,
       "throttle_ratio": 0.602,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -11.5
+      "ttft_p99_drift_ms": -11.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2804.8,
+        "std": 516.6,
+        "cv_pct": 18.42,
+        "stability": "high-variance",
+        "runs": [
+          3817.9,
+          4192.3,
+          2621.3,
+          2548.3,
+          2703.0,
+          2680.5,
+          2557.3,
+          2641.8,
+          2612.4,
+          2524.0,
+          2626.8,
+          2534.0,
+          2647.2,
+          2560.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/sustained/result.json
index 04f3b842..38392d34 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx1_suite_F_nvidia_vllm_47f5d58e_54d0e7aa/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 2804.8,
       "throttle_ratio": 0.602,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -11.5
+      "ttft_p99_drift_ms": -11.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2804.8,
+        "std": 516.6,
+        "cv_pct": 18.42,
+        "stability": "high-variance",
+        "runs": [
+          3817.9,
+          4192.3,
+          2621.3,
+          2548.3,
+          2703.0,
+          2680.5,
+          2557.3,
+          2641.8,
+          2612.4,
+          2524.0,
+          2626.8,
+          2534.0,
+          2647.2,
+          2560.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/result.json b/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/result.json
index 832b5d1a..42bded81 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/result.json
@@ -574,7 +574,44 @@
       "sustained_throughput_tokens_per_sec": 184.0,
       "throttle_ratio": 0.816,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -7.8
+      "ttft_p99_drift_ms": -7.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 184.0,
+        "std": 9.5,
+        "cv_pct": 5.15,
+        "stability": "noisy",
+        "runs": [
+          180.1,
+          187.5,
+          193.8,
+          183.5,
+          191.3,
+          175.4,
+          186.1,
+          181.5,
+          189.5,
+          170.3,
+          184.2,
+          178.7,
+          189.4,
+          173.7,
+          187.8,
+          187.0,
+          180.2,
+          188.7,
+          182.7,
+          173.0,
+          202.4,
+          172.3,
+          203.0,
+          174.7,
+          199.2,
+          178.3,
+          165.6,
+          192.4
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 81.8,
diff --git a/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/sustained/result.json
index 698c9eab..491c6168 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx8_suite_B_nvidia_vllm_47f5d58e_de0853fa/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 184.0,
       "throttle_ratio": 0.816,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -7.8
+      "ttft_p99_drift_ms": -7.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 184.0,
+        "std": 9.5,
+        "cv_pct": 5.15,
+        "stability": "noisy",
+        "runs": [
+          180.1,
+          187.5,
+          193.8,
+          183.5,
+          191.3,
+          175.4,
+          186.1,
+          181.5,
+          189.5,
+          170.3,
+          184.2,
+          178.7,
+          189.4,
+          173.7,
+          187.8,
+          187.0,
+          180.2,
+          188.7,
+          182.7,
+          173.0,
+          202.4,
+          172.3,
+          203.0,
+          174.7,
+          199.2,
+          178.3,
+          165.6,
+          192.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/result.json b/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/result.json
index 83525842..269b6f03 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/result.json
@@ -571,7 +571,44 @@
       "sustained_throughput_tokens_per_sec": 569.1,
       "throttle_ratio": 0.845,
       "throttle_onset_minute": 7.0,
-      "ttft_p99_drift_ms": -27.8
+      "ttft_p99_drift_ms": -27.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 569.1,
+        "std": 22.4,
+        "cv_pct": 3.93,
+        "stability": "noisy",
+        "runs": [
+          573.8,
+          571.7,
+          596.7,
+          594.4,
+          584.0,
+          552.0,
+          550.9,
+          538.9,
+          519.3,
+          542.0,
+          614.7,
+          587.1,
+          597.3,
+          593.2,
+          551.6,
+          581.8,
+          573.7,
+          550.5,
+          564.5,
+          571.8,
+          535.8,
+          589.0,
+          565.8,
+          585.7,
+          578.7,
+          564.8,
+          550.8,
+          553.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/sustained/result.json b/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/sustained/result.json
index 3f730cb1..33416153 100644
--- a/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/sustained/result.json
+++ b/results/verified/nvidia_a800_sxm4_80gbx8_suite_G_nvidia_vllm_47f5d58e_d31ba78b/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 569.1,
       "throttle_ratio": 0.845,
       "throttle_onset_minute": 7.0,
-      "ttft_p99_drift_ms": -27.8
+      "ttft_p99_drift_ms": -27.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 569.1,
+        "std": 22.4,
+        "cv_pct": 3.93,
+        "stability": "noisy",
+        "runs": [
+          573.8,
+          571.7,
+          596.7,
+          594.4,
+          584.0,
+          552.0,
+          550.9,
+          538.9,
+          519.3,
+          542.0,
+          614.7,
+          587.1,
+          597.3,
+          593.2,
+          551.6,
+          581.8,
+          573.7,
+          550.5,
+          564.5,
+          571.8,
+          535.8,
+          589.0,
+          565.8,
+          585.7,
+          578.7,
+          564.8,
+          550.8,
+          553.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/result.json
index 82157fb0..0142e7ff 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/result.json
@@ -504,7 +504,44 @@
       "sustained_throughput_tokens_per_sec": 309.9,
       "throttle_ratio": 0.887,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": 5.6
+      "ttft_p99_drift_ms": 5.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 309.9,
+        "std": 9.6,
+        "cv_pct": 3.1,
+        "stability": "noisy",
+        "runs": [
+          316.9,
+          312.1,
+          306.3,
+          311.7,
+          323.5,
+          307.9,
+          290.3,
+          315.2,
+          308.3,
+          304.1,
+          316.4,
+          309.6,
+          296.7,
+          317.6,
+          311.5,
+          306.4,
+          312.3,
+          321.9,
+          308.5,
+          289.1,
+          323.1,
+          300.0,
+          312.4,
+          306.1,
+          322.1,
+          305.8,
+          296.1,
+          326.1
+        ]
+      }
     },
     "burst": {
       "sla_ttft_ms": 500,
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/sustained/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/sustained/result.json
index 8a59eb57..f4650348 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_A_nvidia_vllm_47f5d58e_e95e2caa/sustained/result.json
@@ -407,7 +407,44 @@
       "sustained_throughput_tokens_per_sec": 309.9,
       "throttle_ratio": 0.887,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": 5.6
+      "ttft_p99_drift_ms": 5.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 309.9,
+        "std": 9.6,
+        "cv_pct": 3.1,
+        "stability": "noisy",
+        "runs": [
+          316.9,
+          312.1,
+          306.3,
+          311.7,
+          323.5,
+          307.9,
+          290.3,
+          315.2,
+          308.3,
+          304.1,
+          316.4,
+          309.6,
+          296.7,
+          317.6,
+          311.5,
+          306.4,
+          312.3,
+          321.9,
+          308.5,
+          289.1,
+          323.1,
+          300.0,
+          312.4,
+          306.1,
+          322.1,
+          305.8,
+          296.1,
+          326.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/result.json
index c3c3e30e..72dbddc6 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/result.json
@@ -366,7 +366,30 @@
       "sustained_throughput_tokens_per_sec": 306.6,
       "throttle_ratio": 0.819,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -306.3
+      "ttft_p99_drift_ms": -306.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 306.6,
+        "std": 14.3,
+        "cv_pct": 4.67,
+        "stability": "noisy",
+        "runs": [
+          267.6,
+          322.3,
+          315.8,
+          306.4,
+          314.6,
+          296.7,
+          326.9,
+          299.4,
+          311.9,
+          301.7,
+          315.6,
+          307.7,
+          306.1,
+          299.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/sustained/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/sustained/result.json
index 97b8ebbf..f1af1137 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/bf16/sustained/result.json
@@ -257,7 +257,30 @@
       "sustained_throughput_tokens_per_sec": 306.6,
       "throttle_ratio": 0.819,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -306.3
+      "ttft_p99_drift_ms": -306.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 306.6,
+        "std": 14.3,
+        "cv_pct": 4.67,
+        "stability": "noisy",
+        "runs": [
+          267.6,
+          322.3,
+          315.8,
+          306.4,
+          314.6,
+          296.7,
+          326.9,
+          299.4,
+          311.9,
+          301.7,
+          315.6,
+          307.7,
+          306.1,
+          299.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/result.json
index c5467b23..b9121bec 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/result.json
@@ -366,7 +366,30 @@
       "sustained_throughput_tokens_per_sec": 472.4,
       "throttle_ratio": 0.918,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -489.7
+      "ttft_p99_drift_ms": -489.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 472.4,
+        "std": 11.9,
+        "cv_pct": 2.51,
+        "stability": "stable",
+        "runs": [
+          455.4,
+          475.3,
+          464.4,
+          488.4,
+          476.5,
+          479.4,
+          465.6,
+          471.6,
+          462.5,
+          468.5,
+          496.3,
+          463.9,
+          460.1,
+          486.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/sustained/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/sustained/result.json
index 5e0a930f..687beb9b 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/fp8/sustained/result.json
@@ -257,7 +257,30 @@
       "sustained_throughput_tokens_per_sec": 472.4,
       "throttle_ratio": 0.918,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -489.7
+      "ttft_p99_drift_ms": -489.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 472.4,
+        "std": 11.9,
+        "cv_pct": 2.51,
+        "stability": "stable",
+        "runs": [
+          455.4,
+          475.3,
+          464.4,
+          488.4,
+          476.5,
+          479.4,
+          465.6,
+          471.6,
+          462.5,
+          468.5,
+          496.3,
+          463.9,
+          460.1,
+          486.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/result.json
index 74c8204c..3044937c 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/result.json
@@ -366,7 +366,30 @@
       "sustained_throughput_tokens_per_sec": 588.9,
       "throttle_ratio": 0.948,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -466.4
+      "ttft_p99_drift_ms": -466.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 588.9,
+        "std": 8.1,
+        "cv_pct": 1.38,
+        "stability": "stable",
+        "runs": [
+          578.0,
+          592.6,
+          590.0,
+          609.6,
+          583.9,
+          593.0,
+          583.9,
+          584.0,
+          584.1,
+          599.7,
+          584.2,
+          584.8,
+          592.3,
+          584.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/sustained/result.json
index 64fff39e..19515355 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w4a16/sustained/result.json
@@ -257,7 +257,30 @@
       "sustained_throughput_tokens_per_sec": 588.9,
       "throttle_ratio": 0.948,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -466.4
+      "ttft_p99_drift_ms": -466.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 588.9,
+        "std": 8.1,
+        "cv_pct": 1.38,
+        "stability": "stable",
+        "runs": [
+          578.0,
+          592.6,
+          590.0,
+          609.6,
+          583.9,
+          593.0,
+          583.9,
+          584.0,
+          584.1,
+          599.7,
+          584.2,
+          584.8,
+          592.3,
+          584.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/result.json
index ffb70f72..8a210904 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/result.json
@@ -366,7 +366,30 @@
       "sustained_throughput_tokens_per_sec": 475.6,
       "throttle_ratio": 0.907,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -484.4
+      "ttft_p99_drift_ms": -484.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 475.6,
+        "std": 12.6,
+        "cv_pct": 2.64,
+        "stability": "stable",
+        "runs": [
+          449.0,
+          494.9,
+          484.5,
+          476.7,
+          460.0,
+          478.3,
+          478.5,
+          476.5,
+          484.4,
+          460.0,
+          478.1,
+          479.4,
+          467.4,
+          490.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/sustained/result.json
index 3f598712..0c81366f 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a16/sustained/result.json
@@ -257,7 +257,30 @@
       "sustained_throughput_tokens_per_sec": 475.6,
       "throttle_ratio": 0.907,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -484.4
+      "ttft_p99_drift_ms": -484.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 475.6,
+        "std": 12.6,
+        "cv_pct": 2.64,
+        "stability": "stable",
+        "runs": [
+          449.0,
+          494.9,
+          484.5,
+          476.7,
+          460.0,
+          478.3,
+          478.5,
+          476.5,
+          484.4,
+          460.0,
+          478.1,
+          479.4,
+          467.4,
+          490.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/result.json
index d367a2b7..db9b3c8a 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/result.json
@@ -366,7 +366,30 @@
       "sustained_throughput_tokens_per_sec": 475.9,
       "throttle_ratio": 0.925,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -57.8
+      "ttft_p99_drift_ms": -57.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 475.9,
+        "std": 9.2,
+        "cv_pct": 1.93,
+        "stability": "stable",
+        "runs": [
+          453.3,
+          472.7,
+          483.2,
+          471.1,
+          490.0,
+          477.4,
+          472.5,
+          471.6,
+          479.5,
+          488.8,
+          475.3,
+          472.2,
+          483.4,
+          471.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/sustained/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/sustained/result.json
index c2f822d9..b71ee9c9 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_C_nvidia_vllm_47f5d58e_4955fbb1/w8a8/sustained/result.json
@@ -257,7 +257,30 @@
       "sustained_throughput_tokens_per_sec": 475.9,
       "throttle_ratio": 0.925,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -57.8
+      "ttft_p99_drift_ms": -57.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 475.9,
+        "std": 9.2,
+        "cv_pct": 1.93,
+        "stability": "stable",
+        "runs": [
+          453.3,
+          472.7,
+          483.2,
+          471.1,
+          490.0,
+          477.4,
+          472.5,
+          471.6,
+          479.5,
+          488.8,
+          475.3,
+          472.2,
+          483.4,
+          471.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/result.json
index b52db8fa..728692e0 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/result.json
@@ -341,7 +341,30 @@
       "sustained_throughput_tokens_per_sec": 2693.3,
       "throttle_ratio": 0.821,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -63.4
+      "ttft_p99_drift_ms": -63.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2693.3,
+        "std": 140.0,
+        "cv_pct": 5.2,
+        "stability": "noisy",
+        "runs": [
+          3023.8,
+          2938.2,
+          2481.2,
+          2558.2,
+          2650.5,
+          2645.4,
+          2618.4,
+          2683.7,
+          2624.8,
+          2660.3,
+          2705.4,
+          2652.5,
+          2764.3,
+          2699.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/sustained/result.json b/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/sustained/result.json
index c26d796f..9b8e14c5 100644
--- a/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_3090x1_suite_F_nvidia_vllm_47f5d58e_faf550ec/sustained/result.json
@@ -257,7 +257,30 @@
       "sustained_throughput_tokens_per_sec": 2693.3,
       "throttle_ratio": 0.821,
       "throttle_onset_minute": 3.0,
-      "ttft_p99_drift_ms": -63.4
+      "ttft_p99_drift_ms": -63.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2693.3,
+        "std": 140.0,
+        "cv_pct": 5.2,
+        "stability": "noisy",
+        "runs": [
+          3023.8,
+          2938.2,
+          2481.2,
+          2558.2,
+          2650.5,
+          2645.4,
+          2618.4,
+          2683.7,
+          2624.8,
+          2660.3,
+          2705.4,
+          2652.5,
+          2764.3,
+          2699.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/result.json
index 7e3068b3..1aebc5c9 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/result.json
@@ -508,7 +508,44 @@
       "sustained_throughput_tokens_per_sec": 400.1,
       "throttle_ratio": 0.936,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -2.6
+      "ttft_p99_drift_ms": -2.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 400.1,
+        "std": 6.3,
+        "cv_pct": 1.58,
+        "stability": "stable",
+        "runs": [
+          406.1,
+          399.7,
+          406.1,
+          387.3,
+          399.9,
+          393.1,
+          403.1,
+          407.1,
+          395.6,
+          399.7,
+          401.3,
+          393.3,
+          407.1,
+          406.0,
+          399.4,
+          402.4,
+          391.0,
+          399.7,
+          384.6,
+          410.9,
+          403.4,
+          398.5,
+          399.7,
+          400.9,
+          393.3,
+          407.0,
+          405.7,
+          399.9
+        ]
+      }
     },
     "burst": {
       "sla_ttft_ms": 500,
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/sustained/result.json
index 8b3f7656..bc017356 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_A_nvidia_vllm_47f5d58e_d6543f77/sustained/result.json
@@ -411,7 +411,44 @@
       "sustained_throughput_tokens_per_sec": 400.1,
       "throttle_ratio": 0.936,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -2.6
+      "ttft_p99_drift_ms": -2.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 400.1,
+        "std": 6.3,
+        "cv_pct": 1.58,
+        "stability": "stable",
+        "runs": [
+          406.1,
+          399.7,
+          406.1,
+          387.3,
+          399.9,
+          393.1,
+          403.1,
+          407.1,
+          395.6,
+          399.7,
+          401.3,
+          393.3,
+          407.1,
+          406.0,
+          399.4,
+          402.4,
+          391.0,
+          399.7,
+          384.6,
+          410.9,
+          403.4,
+          398.5,
+          399.7,
+          400.9,
+          393.3,
+          407.0,
+          405.7,
+          399.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/result.json
index 4bf3f78d..ded7ce76 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/result.json
@@ -370,7 +370,30 @@
       "sustained_throughput_tokens_per_sec": 393.1,
       "throttle_ratio": 0.905,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -186.0
+      "ttft_p99_drift_ms": -186.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 393.1,
+        "std": 11.9,
+        "cv_pct": 3.03,
+        "stability": "noisy",
+        "runs": [
+          368.4,
+          385.6,
+          407.1,
+          388.2,
+          401.4,
+          392.9,
+          407.1,
+          385.4,
+          399.1,
+          373.5,
+          405.2,
+          399.6,
+          391.1,
+          398.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/sustained/result.json
index f6741165..bda3d0bc 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/bf16/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 393.1,
       "throttle_ratio": 0.905,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -186.0
+      "ttft_p99_drift_ms": -186.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 393.1,
+        "std": 11.9,
+        "cv_pct": 3.03,
+        "stability": "noisy",
+        "runs": [
+          368.4,
+          385.6,
+          407.1,
+          388.2,
+          401.4,
+          392.9,
+          407.1,
+          385.4,
+          399.1,
+          373.5,
+          405.2,
+          399.6,
+          391.1,
+          398.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/result.json
index a00e451c..cac12c65 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/result.json
@@ -370,7 +370,30 @@
       "sustained_throughput_tokens_per_sec": 640.7,
       "throttle_ratio": 0.915,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -121.3
+      "ttft_p99_drift_ms": -121.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 640.7,
+        "std": 13.7,
+        "cv_pct": 2.14,
+        "stability": "stable",
+        "runs": [
+          603.2,
+          639.6,
+          630.9,
+          655.0,
+          638.3,
+          638.3,
+          651.3,
+          641.3,
+          647.4,
+          641.4,
+          659.2,
+          629.4,
+          645.2,
+          649.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/sustained/result.json
index e6c613d8..4983b045 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/fp8/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 640.7,
       "throttle_ratio": 0.915,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -121.3
+      "ttft_p99_drift_ms": -121.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 640.7,
+        "std": 13.7,
+        "cv_pct": 2.14,
+        "stability": "stable",
+        "runs": [
+          603.2,
+          639.6,
+          630.9,
+          655.0,
+          638.3,
+          638.3,
+          651.3,
+          641.3,
+          647.4,
+          641.4,
+          659.2,
+          629.4,
+          645.2,
+          649.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/result.json
index 69c78edd..8a8b9f06 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/result.json
@@ -370,7 +370,30 @@
       "sustained_throughput_tokens_per_sec": 854.9,
       "throttle_ratio": 0.919,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -236.8
+      "ttft_p99_drift_ms": -236.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 854.9,
+        "std": 19.1,
+        "cv_pct": 2.23,
+        "stability": "stable",
+        "runs": [
+          814.1,
+          856.6,
+          858.9,
+          849.2,
+          850.8,
+          849.4,
+          876.6,
+          828.0,
+          856.1,
+          852.3,
+          846.7,
+          879.9,
+          864.3,
+          886.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/sustained/result.json
index 015f9cbc..97de99c0 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w4a16/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 854.9,
       "throttle_ratio": 0.919,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -236.8
+      "ttft_p99_drift_ms": -236.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 854.9,
+        "std": 19.1,
+        "cv_pct": 2.23,
+        "stability": "stable",
+        "runs": [
+          814.1,
+          856.6,
+          858.9,
+          849.2,
+          850.8,
+          849.4,
+          876.6,
+          828.0,
+          856.1,
+          852.3,
+          846.7,
+          879.9,
+          864.3,
+          886.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/result.json
index 6f634adb..95316948 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/result.json
@@ -370,7 +370,30 @@
       "sustained_throughput_tokens_per_sec": 628.8,
       "throttle_ratio": 0.916,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -237.0
+      "ttft_p99_drift_ms": -237.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 628.8,
+        "std": 16.5,
+        "cv_pct": 2.63,
+        "stability": "stable",
+        "runs": [
+          593.7,
+          648.2,
+          608.6,
+          648.0,
+          610.4,
+          627.1,
+          619.1,
+          648.0,
+          635.4,
+          623.2,
+          640.1,
+          638.9,
+          625.9,
+          636.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/sustained/result.json
index 3f281590..708866f5 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a16/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 628.8,
       "throttle_ratio": 0.916,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -237.0
+      "ttft_p99_drift_ms": -237.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 628.8,
+        "std": 16.5,
+        "cv_pct": 2.63,
+        "stability": "stable",
+        "runs": [
+          593.7,
+          648.2,
+          608.6,
+          648.0,
+          610.4,
+          627.1,
+          619.1,
+          648.0,
+          635.4,
+          623.2,
+          640.1,
+          638.9,
+          625.9,
+          636.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/result.json
index 814021f9..05ca431a 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/result.json
@@ -370,7 +370,30 @@
       "sustained_throughput_tokens_per_sec": 567.7,
       "throttle_ratio": 0.927,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -121.8
+      "ttft_p99_drift_ms": -121.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 567.7,
+        "std": 10.8,
+        "cv_pct": 1.9,
+        "stability": "stable",
+        "runs": [
+          540.8,
+          556.4,
+          564.4,
+          577.5,
+          563.1,
+          568.6,
+          569.2,
+          569.7,
+          569.3,
+          570.9,
+          574.7,
+          583.3,
+          559.2,
+          580.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/sustained/result.json
index 42d3a726..a1dcd617 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_C_nvidia_vllm_47f5d58e_b59b0798/w8a8/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 567.7,
       "throttle_ratio": 0.927,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -121.8
+      "ttft_p99_drift_ms": -121.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 567.7,
+        "std": 10.8,
+        "cv_pct": 1.9,
+        "stability": "stable",
+        "runs": [
+          540.8,
+          556.4,
+          564.4,
+          577.5,
+          563.1,
+          568.6,
+          569.2,
+          569.7,
+          569.3,
+          570.9,
+          574.7,
+          583.3,
+          559.2,
+          580.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/result.json
index b8776e99..252e1be8 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/result.json
@@ -345,7 +345,30 @@
       "sustained_throughput_tokens_per_sec": 5995.2,
       "throttle_ratio": 0.901,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -25.3
+      "ttft_p99_drift_ms": -25.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 5995.2,
+        "std": 266.2,
+        "cv_pct": 4.44,
+        "stability": "noisy",
+        "runs": [
+          6220.0,
+          6244.2,
+          6209.5,
+          6276.3,
+          6214.1,
+          6252.7,
+          6249.8,
+          5974.5,
+          5679.0,
+          5654.4,
+          5674.7,
+          5677.9,
+          5730.7,
+          5874.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/sustained/result.json
index 3098b1c7..df53813f 100644
--- a/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090_dx1_suite_F_nvidia_vllm_47f5d58e_06662a14/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 5995.2,
       "throttle_ratio": 0.901,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -25.3
+      "ttft_p99_drift_ms": -25.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 5995.2,
+        "std": 266.2,
+        "cv_pct": 4.44,
+        "stability": "noisy",
+        "runs": [
+          6220.0,
+          6244.2,
+          6209.5,
+          6276.3,
+          6214.1,
+          6252.7,
+          6249.8,
+          5974.5,
+          5679.0,
+          5654.4,
+          5674.7,
+          5677.9,
+          5730.7,
+          5874.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/result.json
index fb016a51..16fdc40b 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/result.json
@@ -511,7 +511,44 @@
       "sustained_throughput_tokens_per_sec": 339.8,
       "throttle_ratio": 0.91,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -3.1
+      "ttft_p99_drift_ms": -3.1,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 339.8,
+        "std": 8.8,
+        "cv_pct": 2.58,
+        "stability": "stable",
+        "runs": [
+          334.4,
+          352.3,
+          338.5,
+          327.8,
+          339.6,
+          356.3,
+          339.2,
+          348.1,
+          339.9,
+          348.3,
+          339.3,
+          339.5,
+          334.5,
+          339.3,
+          336.5,
+          352.2,
+          329.1,
+          343.0,
+          332.5,
+          340.7,
+          324.4,
+          331.1,
+          356.6,
+          331.0,
+          331.0,
+          351.4,
+          332.6,
+          344.9
+        ]
+      }
     },
     "burst": {
       "sla_ttft_ms": 500,
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/sustained/result.json
index 73561592..c96c9708 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_A_nvidia_vllm_47f5d58e_675e325e/sustained/result.json
@@ -411,7 +411,44 @@
       "sustained_throughput_tokens_per_sec": 339.8,
       "throttle_ratio": 0.91,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -3.1
+      "ttft_p99_drift_ms": -3.1,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 339.8,
+        "std": 8.8,
+        "cv_pct": 2.58,
+        "stability": "stable",
+        "runs": [
+          334.4,
+          352.3,
+          338.5,
+          327.8,
+          339.6,
+          356.3,
+          339.2,
+          348.1,
+          339.9,
+          348.3,
+          339.3,
+          339.5,
+          334.5,
+          339.3,
+          336.5,
+          352.2,
+          329.1,
+          343.0,
+          332.5,
+          340.7,
+          324.4,
+          331.1,
+          356.6,
+          331.0,
+          331.0,
+          351.4,
+          332.6,
+          344.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/result.json
index c97e8e54..a3591049 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/result.json
@@ -374,7 +374,30 @@
       "sustained_throughput_tokens_per_sec": 334.4,
       "throttle_ratio": 0.867,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -349.6
+      "ttft_p99_drift_ms": -349.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 334.4,
+        "std": 11.9,
+        "cv_pct": 3.57,
+        "stability": "noisy",
+        "runs": [
+          304.8,
+          345.4,
+          326.5,
+          351.5,
+          327.2,
+          338.0,
+          326.8,
+          337.5,
+          339.0,
+          343.2,
+          332.3,
+          348.6,
+          326.4,
+          334.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/sustained/result.json
index c5116404..e34a05c5 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/bf16/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 334.4,
       "throttle_ratio": 0.867,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -349.6
+      "ttft_p99_drift_ms": -349.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 334.4,
+        "std": 11.9,
+        "cv_pct": 3.57,
+        "stability": "noisy",
+        "runs": [
+          304.8,
+          345.4,
+          326.5,
+          351.5,
+          327.2,
+          338.0,
+          326.8,
+          337.5,
+          339.0,
+          343.2,
+          332.3,
+          348.6,
+          326.4,
+          334.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/result.json
index 2d419279..c619c421 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/result.json
@@ -374,7 +374,30 @@
       "sustained_throughput_tokens_per_sec": 472.1,
       "throttle_ratio": 0.907,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -85.0
+      "ttft_p99_drift_ms": -85.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 472.1,
+        "std": 11.5,
+        "cv_pct": 2.44,
+        "stability": "stable",
+        "runs": [
+          445.8,
+          458.1,
+          469.9,
+          469.6,
+          465.7,
+          477.8,
+          479.1,
+          487.9,
+          471.7,
+          491.4,
+          465.6,
+          476.4,
+          473.9,
+          476.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/sustained/result.json
index 93a2b9af..8f68cc7a 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/fp8/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 472.1,
       "throttle_ratio": 0.907,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -85.0
+      "ttft_p99_drift_ms": -85.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 472.1,
+        "std": 11.5,
+        "cv_pct": 2.44,
+        "stability": "stable",
+        "runs": [
+          445.8,
+          458.1,
+          469.9,
+          469.6,
+          465.7,
+          477.8,
+          479.1,
+          487.9,
+          471.7,
+          491.4,
+          465.6,
+          476.4,
+          473.9,
+          476.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/result.json
index 0620dcb8..99318739 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/result.json
@@ -374,7 +374,30 @@
       "sustained_throughput_tokens_per_sec": 606.2,
       "throttle_ratio": 0.916,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -238.3
+      "ttft_p99_drift_ms": -238.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 606.2,
+        "std": 15.7,
+        "cv_pct": 2.59,
+        "stability": "stable",
+        "runs": [
+          584.6,
+          597.4,
+          624.2,
+          591.2,
+          638.4,
+          590.3,
+          598.1,
+          616.0,
+          601.1,
+          605.8,
+          616.5,
+          591.5,
+          624.0,
+          607.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/sustained/result.json
index f32285f8..a18a7e21 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w4a16/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 606.2,
       "throttle_ratio": 0.916,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -238.3
+      "ttft_p99_drift_ms": -238.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 606.2,
+        "std": 15.7,
+        "cv_pct": 2.59,
+        "stability": "stable",
+        "runs": [
+          584.6,
+          597.4,
+          624.2,
+          591.2,
+          638.4,
+          590.3,
+          598.1,
+          616.0,
+          601.1,
+          605.8,
+          616.5,
+          591.5,
+          624.0,
+          607.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/result.json
index ef22615a..2b98e86b 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/result.json
@@ -374,7 +374,30 @@
       "sustained_throughput_tokens_per_sec": 506.0,
       "throttle_ratio": 0.938,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -216.0
+      "ttft_p99_drift_ms": -216.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 506.0,
+        "std": 9.3,
+        "cv_pct": 1.84,
+        "stability": "stable",
+        "runs": [
+          488.0,
+          512.2,
+          495.5,
+          515.4,
+          501.6,
+          507.0,
+          508.3,
+          517.7,
+          509.8,
+          505.8,
+          507.7,
+          499.2,
+          520.5,
+          494.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/sustained/result.json
index 8e45b655..539c6bba 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a16/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 506.0,
       "throttle_ratio": 0.938,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -216.0
+      "ttft_p99_drift_ms": -216.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 506.0,
+        "std": 9.3,
+        "cv_pct": 1.84,
+        "stability": "stable",
+        "runs": [
+          488.0,
+          512.2,
+          495.5,
+          515.4,
+          501.6,
+          507.0,
+          508.3,
+          517.7,
+          509.8,
+          505.8,
+          507.7,
+          499.2,
+          520.5,
+          494.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/result.json
index 2f01c7e8..128bca26 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/result.json
@@ -374,7 +374,30 @@
       "sustained_throughput_tokens_per_sec": 438.9,
       "throttle_ratio": 0.931,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -133.3
+      "ttft_p99_drift_ms": -133.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 438.9,
+        "std": 10.7,
+        "cv_pct": 2.43,
+        "stability": "stable",
+        "runs": [
+          419.8,
+          434.5,
+          439.4,
+          447.0,
+          426.1,
+          445.6,
+          447.0,
+          434.7,
+          441.1,
+          448.9,
+          418.8,
+          448.8,
+          449.6,
+          443.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/sustained/result.json
index f0f86a0a..8f9aa498 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_C_nvidia_vllm_47f5d58e_6d7e1d48/w8a8/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 438.9,
       "throttle_ratio": 0.931,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -133.3
+      "ttft_p99_drift_ms": -133.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 438.9,
+        "std": 10.7,
+        "cv_pct": 2.43,
+        "stability": "stable",
+        "runs": [
+          419.8,
+          434.5,
+          439.4,
+          447.0,
+          426.1,
+          445.6,
+          447.0,
+          434.7,
+          441.1,
+          448.9,
+          418.8,
+          448.8,
+          449.6,
+          443.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/result.json
index 8b04786a..bcc41c80 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/result.json
@@ -348,7 +348,30 @@
       "sustained_throughput_tokens_per_sec": 1698.1,
       "throttle_ratio": 0.922,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -105.2
+      "ttft_p99_drift_ms": -105.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1698.1,
+        "std": 39.2,
+        "cv_pct": 2.31,
+        "stability": "stable",
+        "runs": [
+          1715.3,
+          1751.9,
+          1713.1,
+          1671.7,
+          1724.1,
+          1660.5,
+          1687.0,
+          1622.5,
+          1658.9,
+          1760.4,
+          1680.1,
+          1676.2,
+          1735.0,
+          1716.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/sustained/result.json
index db87a8e7..d1779d30 100644
--- a/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x1_suite_F_nvidia_vllm_47f5d58e_b228454f/sustained/result.json
@@ -261,7 +261,30 @@
       "sustained_throughput_tokens_per_sec": 1698.1,
       "throttle_ratio": 0.922,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -105.2
+      "ttft_p99_drift_ms": -105.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1698.1,
+        "std": 39.2,
+        "cv_pct": 2.31,
+        "stability": "stable",
+        "runs": [
+          1715.3,
+          1751.9,
+          1713.1,
+          1671.7,
+          1724.1,
+          1660.5,
+          1687.0,
+          1622.5,
+          1658.9,
+          1760.4,
+          1680.1,
+          1676.2,
+          1735.0,
+          1716.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/result.json b/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/result.json
index af5e0a5c..1c0fe3d4 100644
--- a/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/result.json
@@ -553,7 +553,44 @@
       "sustained_throughput_tokens_per_sec": 104.5,
       "throttle_ratio": 0.761,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 92.3
+      "ttft_p99_drift_ms": 92.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 104.5,
+        "std": 7.2,
+        "cv_pct": 6.86,
+        "stability": "noisy",
+        "runs": [
+          96.8,
+          106.6,
+          106.6,
+          100.9,
+          109.9,
+          105.4,
+          98.3,
+          105.1,
+          103.0,
+          113.9,
+          101.1,
+          95.9,
+          108.9,
+          110.1,
+          94.4,
+          102.9,
+          111.3,
+          103.9,
+          118.0,
+          89.8,
+          114.0,
+          101.2,
+          95.2,
+          108.8,
+          115.3,
+          94.1,
+          109.8,
+          104.4
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 277.19,
diff --git a/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/sustained/result.json
index cb86a2ad..249553cd 100644
--- a/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x8_suite_B_nvidia_vllm_47f5d58e_cfd0bdc8/sustained/result.json
@@ -451,7 +451,44 @@
       "sustained_throughput_tokens_per_sec": 104.5,
       "throttle_ratio": 0.761,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 92.3
+      "ttft_p99_drift_ms": 92.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 104.5,
+        "std": 7.2,
+        "cv_pct": 6.86,
+        "stability": "noisy",
+        "runs": [
+          96.8,
+          106.6,
+          106.6,
+          100.9,
+          109.9,
+          105.4,
+          98.3,
+          105.1,
+          103.0,
+          113.9,
+          101.1,
+          95.9,
+          108.9,
+          110.1,
+          94.4,
+          102.9,
+          111.3,
+          103.9,
+          118.0,
+          89.8,
+          114.0,
+          101.2,
+          95.2,
+          108.8,
+          115.3,
+          94.1,
+          109.8,
+          104.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/result.json b/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/result.json
index f57c2d05..678c24fd 100644
--- a/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/result.json
@@ -550,7 +550,44 @@
       "sustained_throughput_tokens_per_sec": 325.5,
       "throttle_ratio": 0.886,
       "throttle_onset_minute": 5.0,
-      "ttft_p99_drift_ms": -19.3
+      "ttft_p99_drift_ms": -19.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 325.5,
+        "std": 9.7,
+        "cv_pct": 2.99,
+        "stability": "stable",
+        "runs": [
+          348.2,
+          317.8,
+          340.5,
+          312.2,
+          342.5,
+          323.8,
+          311.0,
+          329.7,
+          314.2,
+          308.6,
+          338.2,
+          323.2,
+          326.9,
+          326.3,
+          324.1,
+          329.2,
+          329.8,
+          323.9,
+          333.7,
+          318.7,
+          329.3,
+          316.3,
+          317.7,
+          334.9,
+          329.4,
+          322.7,
+          321.0,
+          319.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/sustained/result.json b/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/sustained/result.json
index 433f8f6f..405ca5df 100644
--- a/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_4090x8_suite_G_nvidia_vllm_47f5d58e_a4179ecc/sustained/result.json
@@ -451,7 +451,44 @@
       "sustained_throughput_tokens_per_sec": 325.5,
       "throttle_ratio": 0.886,
       "throttle_onset_minute": 5.0,
-      "ttft_p99_drift_ms": -19.3
+      "ttft_p99_drift_ms": -19.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 325.5,
+        "std": 9.7,
+        "cv_pct": 2.99,
+        "stability": "stable",
+        "runs": [
+          348.2,
+          317.8,
+          340.5,
+          312.2,
+          342.5,
+          323.8,
+          311.0,
+          329.7,
+          314.2,
+          308.6,
+          338.2,
+          323.2,
+          326.9,
+          326.3,
+          324.1,
+          329.2,
+          329.8,
+          323.9,
+          333.7,
+          318.7,
+          329.3,
+          316.3,
+          317.7,
+          334.9,
+          329.4,
+          322.7,
+          321.0,
+          319.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/result.json
index 1308f73d..11baab5a 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/result.json
@@ -522,7 +522,44 @@
       "sustained_throughput_tokens_per_sec": 707.5,
       "throttle_ratio": 0.948,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -33.3
+      "ttft_p99_drift_ms": -33.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 707.5,
+        "std": 9.9,
+        "cv_pct": 1.4,
+        "stability": "stable",
+        "runs": [
+          700.3,
+          705.2,
+          720.8,
+          690.5,
+          705.3,
+          718.8,
+          706.6,
+          711.1,
+          700.4,
+          708.3,
+          720.0,
+          710.9,
+          691.8,
+          699.3,
+          726.0,
+          702.1,
+          708.6,
+          701.3,
+          722.9,
+          700.6,
+          714.3,
+          701.7,
+          695.2,
+          728.2,
+          699.1,
+          705.2,
+          708.6,
+          707.1
+        ]
+      }
     },
     "burst": {
       "sla_ttft_ms": 500,
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/sustained/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/sustained/result.json
index cf184255..62c1d274 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_A_nvidia_vllm_47f5d58e_b8f8ed0f/sustained/result.json
@@ -422,7 +422,44 @@
       "sustained_throughput_tokens_per_sec": 707.5,
       "throttle_ratio": 0.948,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -33.3
+      "ttft_p99_drift_ms": -33.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 707.5,
+        "std": 9.9,
+        "cv_pct": 1.4,
+        "stability": "stable",
+        "runs": [
+          700.3,
+          705.2,
+          720.8,
+          690.5,
+          705.3,
+          718.8,
+          706.6,
+          711.1,
+          700.4,
+          708.3,
+          720.0,
+          710.9,
+          691.8,
+          699.3,
+          726.0,
+          702.1,
+          708.6,
+          701.3,
+          722.9,
+          700.6,
+          714.3,
+          701.7,
+          695.2,
+          728.2,
+          699.1,
+          705.2,
+          708.6,
+          707.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/result.json
index fb307d1f..cc31153f 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/result.json
@@ -425,7 +425,30 @@
       "sustained_throughput_tokens_per_sec": 676.2,
       "throttle_ratio": 0.429,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -29850.4
+      "ttft_p99_drift_ms": -29850.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 676.2,
+        "std": 105.7,
+        "cv_pct": 15.64,
+        "stability": "high-variance",
+        "runs": [
+          310.3,
+          701.9,
+          691.2,
+          717.0,
+          697.2,
+          705.7,
+          708.4,
+          704.3,
+          690.8,
+          713.3,
+          698.7,
+          694.2,
+          710.3,
+          723.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/sustained/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/sustained/result.json
index 9d9b9a61..7a3c03f4 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/bf16/sustained/result.json
@@ -312,7 +312,30 @@
       "sustained_throughput_tokens_per_sec": 676.2,
       "throttle_ratio": 0.429,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -29850.4
+      "ttft_p99_drift_ms": -29850.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 676.2,
+        "std": 105.7,
+        "cv_pct": 15.64,
+        "stability": "high-variance",
+        "runs": [
+          310.3,
+          701.9,
+          691.2,
+          717.0,
+          697.2,
+          705.7,
+          708.4,
+          704.3,
+          690.8,
+          713.3,
+          698.7,
+          694.2,
+          710.3,
+          723.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/result.json
index eb2298d0..3cb556fc 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/result.json
@@ -425,7 +425,30 @@
       "sustained_throughput_tokens_per_sec": 1381.4,
       "throttle_ratio": 0.439,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -30389.8
+      "ttft_p99_drift_ms": -30389.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1381.4,
+        "std": 213.7,
+        "cv_pct": 15.47,
+        "stability": "high-variance",
+        "runs": [
+          641.1,
+          1449.1,
+          1415.5,
+          1411.0,
+          1429.2,
+          1424.7,
+          1453.5,
+          1454.5,
+          1422.3,
+          1435.2,
+          1456.1,
+          1440.6,
+          1447.5,
+          1459.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/sustained/result.json
index 22d6cdb7..ca9ac78b 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w4a16/sustained/result.json
@@ -312,7 +312,30 @@
       "sustained_throughput_tokens_per_sec": 1381.4,
       "throttle_ratio": 0.439,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -30389.8
+      "ttft_p99_drift_ms": -30389.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1381.4,
+        "std": 213.7,
+        "cv_pct": 15.47,
+        "stability": "high-variance",
+        "runs": [
+          641.1,
+          1449.1,
+          1415.5,
+          1411.0,
+          1429.2,
+          1424.7,
+          1453.5,
+          1454.5,
+          1422.3,
+          1435.2,
+          1456.1,
+          1440.6,
+          1447.5,
+          1459.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/result.json
index bac4cb5a..770f027f 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/result.json
@@ -425,7 +425,30 @@
       "sustained_throughput_tokens_per_sec": 1148.7,
       "throttle_ratio": 0.36,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -35987.3
+      "ttft_p99_drift_ms": -35987.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1148.7,
+        "std": 204.2,
+        "cv_pct": 17.78,
+        "stability": "high-variance",
+        "runs": [
+          440.3,
+          1192.8,
+          1217.1,
+          1190.1,
+          1223.8,
+          1197.8,
+          1200.5,
+          1215.8,
+          1201.4,
+          1197.9,
+          1216.1,
+          1187.5,
+          1191.1,
+          1209.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/sustained/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/sustained/result.json
index 7ae61d5f..15ef4dbb 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_C_nvidia_vllm_47f5d58e_d1baa050/w8a16/sustained/result.json
@@ -312,7 +312,30 @@
       "sustained_throughput_tokens_per_sec": 1148.7,
       "throttle_ratio": 0.36,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -35987.3
+      "ttft_p99_drift_ms": -35987.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1148.7,
+        "std": 204.2,
+        "cv_pct": 17.78,
+        "stability": "high-variance",
+        "runs": [
+          440.3,
+          1192.8,
+          1217.1,
+          1190.1,
+          1223.8,
+          1197.8,
+          1200.5,
+          1215.8,
+          1201.4,
+          1197.9,
+          1216.1,
+          1187.5,
+          1191.1,
+          1209.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/result.json
index e665b3f9..5b6f5ea4 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/result.json
@@ -466,7 +466,44 @@
       "sustained_throughput_tokens_per_sec": 51.5,
       "throttle_ratio": 0.845,
       "throttle_onset_minute": 11.0,
-      "ttft_p99_drift_ms": -35433.6
+      "ttft_p99_drift_ms": -35433.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 51.5,
+        "std": 2.0,
+        "cv_pct": 3.83,
+        "stability": "noisy",
+        "runs": [
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          55.4,
+          51.3,
+          51.2,
+          46.9,
+          55.4,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          55.4,
+          47.0,
+          55.5,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/sustained/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/sustained/result.json
index 1c772358..5d0c5acb 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_D_nvidia_vllm_47f5d58e_e87e6c36/sustained/result.json
@@ -422,7 +422,44 @@
       "sustained_throughput_tokens_per_sec": 51.5,
       "throttle_ratio": 0.845,
       "throttle_onset_minute": 11.0,
-      "ttft_p99_drift_ms": -35433.6
+      "ttft_p99_drift_ms": -35433.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 51.5,
+        "std": 2.0,
+        "cv_pct": 3.83,
+        "stability": "noisy",
+        "runs": [
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          55.4,
+          51.3,
+          51.2,
+          46.9,
+          55.4,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          55.4,
+          47.0,
+          55.5,
+          51.2,
+          51.2,
+          51.2,
+          51.2,
+          51.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/result.json
index cac5f1b6..1cf0d911 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/result.json
@@ -359,7 +359,30 @@
       "sustained_throughput_tokens_per_sec": 3941.2,
       "throttle_ratio": 0.137,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -49045.0
+      "ttft_p99_drift_ms": -49045.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 3941.2,
+        "std": 974.0,
+        "cv_pct": 24.71,
+        "stability": "high-variance",
+        "runs": [
+          683.0,
+          4100.0,
+          4041.5,
+          4995.0,
+          4429.7,
+          4028.5,
+          4021.4,
+          4158.3,
+          3928.9,
+          4175.5,
+          4112.6,
+          4274.6,
+          4203.8,
+          4024.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/sustained/result.json b/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/sustained/result.json
index 7dc690da..45594d91 100644
--- a/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/sustained/result.json
+++ b/results/verified/nvidia_geforce_rtx_5090x1_suite_F_nvidia_vllm_47f5d58e_776d2702/sustained/result.json
@@ -272,7 +272,30 @@
       "sustained_throughput_tokens_per_sec": 3941.2,
       "throttle_ratio": 0.137,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -49045.0
+      "ttft_p99_drift_ms": -49045.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 3941.2,
+        "std": 974.0,
+        "cv_pct": 24.71,
+        "stability": "high-variance",
+        "runs": [
+          683.0,
+          4100.0,
+          4041.5,
+          4995.0,
+          4429.7,
+          4028.5,
+          4021.4,
+          4158.3,
+          3928.9,
+          4175.5,
+          4112.6,
+          4274.6,
+          4203.8,
+          4024.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/result.json b/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/result.json
index 96df1a3f..719e3163 100644
--- a/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/result.json
+++ b/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/result.json
@@ -478,7 +478,44 @@
       "sustained_throughput_tokens_per_sec": 907.1,
       "throttle_ratio": 0.948,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -1.0
+      "ttft_p99_drift_ms": -1.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 907.1,
+        "std": 11.3,
+        "cv_pct": 1.25,
+        "stability": "stable",
+        "runs": [
+          932.2,
+          905.7,
+          909.4,
+          893.4,
+          919.9,
+          906.1,
+          904.7,
+          913.0,
+          911.6,
+          910.9,
+          894.4,
+          912.3,
+          910.6,
+          905.7,
+          910.2,
+          903.0,
+          898.6,
+          911.5,
+          888.3,
+          911.6,
+          900.5,
+          903.7,
+          924.4,
+          883.7,
+          911.4,
+          914.1,
+          922.4,
+          884.7
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/sustained/result.json b/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/sustained/result.json
index f43caa42..325ab779 100644
--- a/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/sustained/result.json
+++ b/results/verified/nvidia_h100_80gb_hbm3x1_suite_A_nvidia_vllm_47f5d58e_831c95a7/sustained/result.json
@@ -377,7 +377,44 @@
       "sustained_throughput_tokens_per_sec": 907.1,
       "throttle_ratio": 0.948,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -1.0
+      "ttft_p99_drift_ms": -1.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 907.1,
+        "std": 11.3,
+        "cv_pct": 1.25,
+        "stability": "stable",
+        "runs": [
+          932.2,
+          905.7,
+          909.4,
+          893.4,
+          919.9,
+          906.1,
+          904.7,
+          913.0,
+          911.6,
+          910.9,
+          894.4,
+          912.3,
+          910.6,
+          905.7,
+          910.2,
+          903.0,
+          898.6,
+          911.5,
+          888.3,
+          911.6,
+          900.5,
+          903.7,
+          924.4,
+          883.7,
+          911.4,
+          914.1,
+          922.4,
+          884.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/result.json b/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/result.json
index f718da39..9e207d46 100644
--- a/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/result.json
+++ b/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/result.json
@@ -421,7 +421,44 @@
       "sustained_throughput_tokens_per_sec": 142.6,
       "throttle_ratio": 0.8,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 327.4
+      "ttft_p99_drift_ms": 327.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 142.6,
+        "std": 13.3,
+        "cv_pct": 9.33,
+        "stability": "high-variance",
+        "runs": [
+          136.5,
+          136.6,
+          136.5,
+          136.5,
+          170.7,
+          136.5,
+          136.5,
+          136.5,
+          136.5,
+          170.7,
+          136.5,
+          136.6,
+          136.6,
+          136.5,
+          170.5,
+          136.5,
+          136.6,
+          136.6,
+          136.5,
+          136.6,
+          170.7,
+          136.5,
+          136.5,
+          136.6,
+          136.5,
+          170.7,
+          136.5,
+          136.6
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/sustained/result.json b/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/sustained/result.json
index dad0da89..5ebc9840 100644
--- a/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/sustained/result.json
+++ b/results/verified/nvidia_h100_80gb_hbm3x1_suite_D_nvidia_vllm_47f5d58e_02748da4/sustained/result.json
@@ -377,7 +377,44 @@
       "sustained_throughput_tokens_per_sec": 142.6,
       "throttle_ratio": 0.8,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 327.4
+      "ttft_p99_drift_ms": 327.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 142.6,
+        "std": 13.3,
+        "cv_pct": 9.33,
+        "stability": "high-variance",
+        "runs": [
+          136.5,
+          136.6,
+          136.5,
+          136.5,
+          170.7,
+          136.5,
+          136.5,
+          136.5,
+          136.5,
+          170.7,
+          136.5,
+          136.6,
+          136.6,
+          136.5,
+          170.5,
+          136.5,
+          136.6,
+          136.6,
+          136.5,
+          136.6,
+          170.7,
+          136.5,
+          136.5,
+          136.6,
+          136.5,
+          170.7,
+          136.5,
+          136.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/result.json b/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/result.json
index 396625d1..3e139b6c 100644
--- a/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/result.json
+++ b/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/result.json
@@ -314,7 +314,30 @@
       "sustained_throughput_tokens_per_sec": 6144.7,
       "throttle_ratio": 0.96,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -19.0
+      "ttft_p99_drift_ms": -19.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 6144.7,
+        "std": 64.6,
+        "cv_pct": 1.05,
+        "stability": "stable",
+        "runs": [
+          6336.2,
+          6162.6,
+          6134.7,
+          6123.2,
+          6087.1,
+          6131.9,
+          6123.4,
+          6204.8,
+          6085.4,
+          6082.5,
+          6153.4,
+          6160.7,
+          6110.4,
+          6129.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/sustained/result.json b/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/sustained/result.json
index 9d299d69..007da95e 100644
--- a/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/sustained/result.json
+++ b/results/verified/nvidia_h100_80gb_hbm3x1_suite_F_nvidia_vllm_47f5d58e_2c0b7beb/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 6144.7,
       "throttle_ratio": 0.96,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -19.0
+      "ttft_p99_drift_ms": -19.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 6144.7,
+        "std": 64.6,
+        "cv_pct": 1.05,
+        "stability": "stable",
+        "runs": [
+          6336.2,
+          6162.6,
+          6134.7,
+          6123.2,
+          6087.1,
+          6131.9,
+          6123.4,
+          6204.8,
+          6085.4,
+          6082.5,
+          6153.4,
+          6160.7,
+          6110.4,
+          6129.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/result.json b/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/result.json
index b67c6d7d..6a5294d5 100644
--- a/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/result.json
+++ b/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/result.json
@@ -526,7 +526,44 @@
       "sustained_throughput_tokens_per_sec": 709.2,
       "throttle_ratio": 0.791,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": 45.9
+      "ttft_p99_drift_ms": 45.9,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 709.2,
+        "std": 40.9,
+        "cv_pct": 5.77,
+        "stability": "noisy",
+        "runs": [
+          851.5,
+          842.9,
+          677.1,
+          719.3,
+          673.5,
+          699.1,
+          702.7,
+          686.1,
+          706.4,
+          676.1,
+          719.9,
+          690.8,
+          711.4,
+          703.6,
+          704.7,
+          700.4,
+          699.5,
+          694.9,
+          707.4,
+          674.7,
+          705.4,
+          705.4,
+          713.4,
+          690.6,
+          696.4,
+          701.8,
+          696.1,
+          705.6
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/sustained/result.json b/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/sustained/result.json
index 412e3363..bd40f8e2 100644
--- a/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_A_nvidia_vllm_47f5d58e_29b2ec38/sustained/result.json
@@ -428,7 +428,44 @@
       "sustained_throughput_tokens_per_sec": 709.2,
       "throttle_ratio": 0.791,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": 45.9
+      "ttft_p99_drift_ms": 45.9,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 709.2,
+        "std": 40.9,
+        "cv_pct": 5.77,
+        "stability": "noisy",
+        "runs": [
+          851.5,
+          842.9,
+          677.1,
+          719.3,
+          673.5,
+          699.1,
+          702.7,
+          686.1,
+          706.4,
+          676.1,
+          719.9,
+          690.8,
+          711.4,
+          703.6,
+          704.7,
+          700.4,
+          699.5,
+          694.9,
+          707.4,
+          674.7,
+          705.4,
+          705.4,
+          713.4,
+          690.6,
+          696.4,
+          701.8,
+          696.1,
+          705.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/result.json
index af08ca34..43b139da 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/result.json
@@ -457,7 +457,30 @@
       "sustained_throughput_tokens_per_sec": 709.4,
       "throttle_ratio": 0.868,
       "throttle_onset_minute": 13.0,
-      "ttft_p99_drift_ms": -149.4
+      "ttft_p99_drift_ms": -149.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 709.4,
+        "std": 30.8,
+        "cv_pct": 4.34,
+        "stability": "noisy",
+        "runs": [
+          709.5,
+          703.8,
+          729.3,
+          727.1,
+          713.1,
+          728.9,
+          721.0,
+          715.9,
+          726.3,
+          734.4,
+          716.6,
+          727.1,
+          641.9,
+          637.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/sustained/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/sustained/result.json
index dc96dc06..5d2bc3cc 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/bf16/sustained/result.json
@@ -348,7 +348,30 @@
       "sustained_throughput_tokens_per_sec": 709.4,
       "throttle_ratio": 0.868,
       "throttle_onset_minute": 13.0,
-      "ttft_p99_drift_ms": -149.4
+      "ttft_p99_drift_ms": -149.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 709.4,
+        "std": 30.8,
+        "cv_pct": 4.34,
+        "stability": "noisy",
+        "runs": [
+          709.5,
+          703.8,
+          729.3,
+          727.1,
+          713.1,
+          728.9,
+          721.0,
+          715.9,
+          726.3,
+          734.4,
+          716.6,
+          727.1,
+          641.9,
+          637.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/result.json
index 2d2ab534..3784d212 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/result.json
@@ -457,7 +457,30 @@
       "sustained_throughput_tokens_per_sec": 713.4,
       "throttle_ratio": 0.888,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -69.4
+      "ttft_p99_drift_ms": -69.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 713.4,
+        "std": 21.2,
+        "cv_pct": 2.97,
+        "stability": "stable",
+        "runs": [
+          736.7,
+          762.2,
+          721.6,
+          676.8,
+          719.7,
+          711.0,
+          715.7,
+          695.0,
+          712.8,
+          702.3,
+          727.1,
+          685.7,
+          708.4,
+          712.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/sustained/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/sustained/result.json
index f1c1cbfb..785ddd38 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/fp8/sustained/result.json
@@ -348,7 +348,30 @@
       "sustained_throughput_tokens_per_sec": 713.4,
       "throttle_ratio": 0.888,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -69.4
+      "ttft_p99_drift_ms": -69.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 713.4,
+        "std": 21.2,
+        "cv_pct": 2.97,
+        "stability": "stable",
+        "runs": [
+          736.7,
+          762.2,
+          721.6,
+          676.8,
+          719.7,
+          711.0,
+          715.7,
+          695.0,
+          712.8,
+          702.3,
+          727.1,
+          685.7,
+          708.4,
+          712.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/result.json
index fd1ef782..73d2749d 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/result.json
@@ -457,7 +457,30 @@
       "sustained_throughput_tokens_per_sec": 649.9,
       "throttle_ratio": 0.886,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -146.9
+      "ttft_p99_drift_ms": -146.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 649.9,
+        "std": 21.8,
+        "cv_pct": 3.35,
+        "stability": "noisy",
+        "runs": [
+          713.6,
+          677.3,
+          649.8,
+          636.1,
+          652.4,
+          646.2,
+          633.3,
+          649.0,
+          638.9,
+          655.1,
+          642.4,
+          636.4,
+          632.1,
+          636.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/sustained/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/sustained/result.json
index 6caffea6..9705f2f2 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w4a16/sustained/result.json
@@ -348,7 +348,30 @@
       "sustained_throughput_tokens_per_sec": 649.9,
       "throttle_ratio": 0.886,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -146.9
+      "ttft_p99_drift_ms": -146.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 649.9,
+        "std": 21.8,
+        "cv_pct": 3.35,
+        "stability": "noisy",
+        "runs": [
+          713.6,
+          677.3,
+          649.8,
+          636.1,
+          652.4,
+          646.2,
+          633.3,
+          649.0,
+          638.9,
+          655.1,
+          642.4,
+          636.4,
+          632.1,
+          636.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/result.json
index 488b494f..51797ff3 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/result.json
@@ -457,7 +457,30 @@
       "sustained_throughput_tokens_per_sec": 708.2,
       "throttle_ratio": 0.878,
       "throttle_onset_minute": 11.0,
-      "ttft_p99_drift_ms": -145.8
+      "ttft_p99_drift_ms": -145.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 708.2,
+        "std": 34.8,
+        "cv_pct": 4.91,
+        "stability": "noisy",
+        "runs": [
+          710.4,
+          720.8,
+          741.0,
+          742.8,
+          739.9,
+          728.3,
+          734.6,
+          723.4,
+          737.2,
+          704.2,
+          664.2,
+          654.3,
+          662.2,
+          652.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/sustained/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/sustained/result.json
index 801e6a43..4ef6e3cb 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a16/sustained/result.json
@@ -348,7 +348,30 @@
       "sustained_throughput_tokens_per_sec": 708.2,
       "throttle_ratio": 0.878,
       "throttle_onset_minute": 11.0,
-      "ttft_p99_drift_ms": -145.8
+      "ttft_p99_drift_ms": -145.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 708.2,
+        "std": 34.8,
+        "cv_pct": 4.91,
+        "stability": "noisy",
+        "runs": [
+          710.4,
+          720.8,
+          741.0,
+          742.8,
+          739.9,
+          728.3,
+          734.6,
+          723.4,
+          737.2,
+          704.2,
+          664.2,
+          654.3,
+          662.2,
+          652.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/result.json
index 980a521e..13d327f9 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/result.json
@@ -457,7 +457,30 @@
       "sustained_throughput_tokens_per_sec": 694.7,
       "throttle_ratio": 0.932,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -100.3
+      "ttft_p99_drift_ms": -100.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 694.7,
+        "std": 12.1,
+        "cv_pct": 1.74,
+        "stability": "stable",
+        "runs": [
+          730.5,
+          688.0,
+          700.2,
+          690.9,
+          686.4,
+          704.0,
+          690.4,
+          691.5,
+          700.7,
+          690.2,
+          695.2,
+          680.6,
+          692.2,
+          685.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/sustained/result.json b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/sustained/result.json
index 736cc3f8..0d28b31f 100644
--- a/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_C_nvidia_vllm_47f5d58e_f07c60f8/w8a8/sustained/result.json
@@ -348,7 +348,30 @@
       "sustained_throughput_tokens_per_sec": 694.7,
       "throttle_ratio": 0.932,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -100.3
+      "ttft_p99_drift_ms": -100.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 694.7,
+        "std": 12.1,
+        "cv_pct": 1.74,
+        "stability": "stable",
+        "runs": [
+          730.5,
+          688.0,
+          700.2,
+          690.9,
+          686.4,
+          704.0,
+          690.4,
+          691.5,
+          700.7,
+          690.2,
+          695.2,
+          680.6,
+          692.2,
+          685.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/result.json b/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/result.json
index d36e3809..023009e0 100644
--- a/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/result.json
+++ b/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/result.json
@@ -471,7 +471,44 @@
       "sustained_throughput_tokens_per_sec": 132.9,
       "throttle_ratio": 0.6,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -55.8
+      "ttft_p99_drift_ms": -55.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 132.9,
+        "std": 14.2,
+        "cv_pct": 10.69,
+        "stability": "high-variance",
+        "runs": [
+          136.5,
+          136.6,
+          136.4,
+          136.5,
+          136.6,
+          136.6,
+          136.4,
+          136.5,
+          136.5,
+          136.6,
+          136.6,
+          136.5,
+          170.7,
+          136.5,
+          136.6,
+          136.5,
+          136.5,
+          136.5,
+          136.6,
+          136.5,
+          136.6,
+          136.5,
+          136.5,
+          136.5,
+          102.4,
+          102.4,
+          102.4,
+          102.4
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/sustained/result.json b/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/sustained/result.json
index 6d8e3212..b27c0571 100644
--- a/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_D_nvidia_vllm_47f5d58e_62a36028/sustained/result.json
@@ -428,7 +428,44 @@
       "sustained_throughput_tokens_per_sec": 132.9,
       "throttle_ratio": 0.6,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -55.8
+      "ttft_p99_drift_ms": -55.8,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 132.9,
+        "std": 14.2,
+        "cv_pct": 10.69,
+        "stability": "high-variance",
+        "runs": [
+          136.5,
+          136.6,
+          136.4,
+          136.5,
+          136.6,
+          136.6,
+          136.4,
+          136.5,
+          136.5,
+          136.6,
+          136.6,
+          136.5,
+          170.7,
+          136.5,
+          136.6,
+          136.5,
+          136.5,
+          136.5,
+          136.6,
+          136.5,
+          136.6,
+          136.5,
+          136.5,
+          136.5,
+          102.4,
+          102.4,
+          102.4,
+          102.4
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/result.json b/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/result.json
index 0f171076..7ffc29f7 100644
--- a/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/result.json
+++ b/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/result.json
@@ -431,7 +431,30 @@
       "sustained_throughput_tokens_per_sec": 1425.4,
       "throttle_ratio": 0.826,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -364.2
+      "ttft_p99_drift_ms": -364.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1425.4,
+        "std": 66.2,
+        "cv_pct": 4.65,
+        "stability": "noisy",
+        "runs": [
+          1306.2,
+          1417.0,
+          1440.4,
+          1411.2,
+          1413.3,
+          1581.2,
+          1523.2,
+          1435.7,
+          1342.6,
+          1396.7,
+          1408.2,
+          1415.5,
+          1436.7,
+          1427.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/sustained/result.json b/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/sustained/result.json
index 3889eef9..0d2a661d 100644
--- a/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/sustained/result.json
+++ b/results/verified/nvidia_h200x1_suite_F_nvidia_vllm_47f5d58e_53471efa/sustained/result.json
@@ -347,7 +347,30 @@
       "sustained_throughput_tokens_per_sec": 1425.4,
       "throttle_ratio": 0.826,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -364.2
+      "ttft_p99_drift_ms": -364.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1425.4,
+        "std": 66.2,
+        "cv_pct": 4.65,
+        "stability": "noisy",
+        "runs": [
+          1306.2,
+          1417.0,
+          1440.4,
+          1411.2,
+          1413.3,
+          1581.2,
+          1523.2,
+          1435.7,
+          1342.6,
+          1396.7,
+          1408.2,
+          1415.5,
+          1436.7,
+          1427.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/result.json b/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/result.json
index 01daefd8..0519bb09 100644
--- a/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/result.json
+++ b/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/result.json
@@ -596,7 +596,44 @@
       "sustained_throughput_tokens_per_sec": 241.2,
       "throttle_ratio": 0.807,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -1.5
+      "ttft_p99_drift_ms": -1.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 241.2,
+        "std": 12.7,
+        "cv_pct": 5.28,
+        "stability": "noisy",
+        "runs": [
+          283.6,
+          281.0,
+          237.6,
+          240.6,
+          232.5,
+          236.4,
+          238.2,
+          242.7,
+          241.5,
+          231.0,
+          244.4,
+          236.3,
+          242.5,
+          235.3,
+          230.9,
+          237.7,
+          243.4,
+          235.9,
+          240.4,
+          232.4,
+          240.7,
+          246.4,
+          232.5,
+          239.3,
+          233.9,
+          235.6,
+          252.2,
+          228.8
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 77.23,
diff --git a/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/sustained/result.json b/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/sustained/result.json
index 3a4cd706..ee0e251c 100644
--- a/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/sustained/result.json
+++ b/results/verified/nvidia_h200x8_suite_B_nvidia_vllm_47f5d58e_b727568e/sustained/result.json
@@ -497,7 +497,44 @@
       "sustained_throughput_tokens_per_sec": 241.2,
       "throttle_ratio": 0.807,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -1.5
+      "ttft_p99_drift_ms": -1.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 241.2,
+        "std": 12.7,
+        "cv_pct": 5.28,
+        "stability": "noisy",
+        "runs": [
+          283.6,
+          281.0,
+          237.6,
+          240.6,
+          232.5,
+          236.4,
+          238.2,
+          242.7,
+          241.5,
+          231.0,
+          244.4,
+          236.3,
+          242.5,
+          235.3,
+          230.9,
+          237.7,
+          243.4,
+          235.9,
+          240.4,
+          232.4,
+          240.7,
+          246.4,
+          232.5,
+          239.3,
+          233.9,
+          235.6,
+          252.2,
+          228.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/result.json b/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/result.json
index b27ea8ea..ffb5d59a 100644
--- a/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/result.json
+++ b/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/result.json
@@ -594,7 +594,44 @@
       "sustained_throughput_tokens_per_sec": 591.5,
       "throttle_ratio": 0.939,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": 12.0
+      "ttft_p99_drift_ms": 12.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 591.5,
+        "std": 9.7,
+        "cv_pct": 1.64,
+        "stability": "stable",
+        "runs": [
+          576.8,
+          613.1,
+          587.8,
+          582.0,
+          595.7,
+          591.5,
+          589.9,
+          585.6,
+          599.3,
+          594.4,
+          596.9,
+          576.0,
+          608.9,
+          579.4,
+          599.9,
+          588.2,
+          607.2,
+          576.5,
+          586.3,
+          582.1,
+          593.9,
+          599.3,
+          597.4,
+          588.5,
+          583.3,
+          598.1,
+          588.0,
+          596.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/sustained/result.json b/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/sustained/result.json
index 08974360..2cf787f8 100644
--- a/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/sustained/result.json
+++ b/results/verified/nvidia_h200x8_suite_G_nvidia_vllm_47f5d58e_7f7a270e/sustained/result.json
@@ -498,7 +498,44 @@
       "sustained_throughput_tokens_per_sec": 591.5,
       "throttle_ratio": 0.939,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": 12.0
+      "ttft_p99_drift_ms": 12.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 591.5,
+        "std": 9.7,
+        "cv_pct": 1.64,
+        "stability": "stable",
+        "runs": [
+          576.8,
+          613.1,
+          587.8,
+          582.0,
+          595.7,
+          591.5,
+          589.9,
+          585.6,
+          599.3,
+          594.4,
+          596.9,
+          576.0,
+          608.9,
+          579.4,
+          599.9,
+          588.2,
+          607.2,
+          576.5,
+          586.3,
+          582.1,
+          593.9,
+          599.3,
+          597.4,
+          588.5,
+          583.3,
+          598.1,
+          588.0,
+          596.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/result.json b/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/result.json
index e6b4b502..71a19486 100644
--- a/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/result.json
@@ -523,7 +523,44 @@
       "sustained_throughput_tokens_per_sec": 486.6,
       "throttle_ratio": 0.918,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": 1.0
+      "ttft_p99_drift_ms": 1.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 486.6,
+        "std": 9.5,
+        "cv_pct": 1.95,
+        "stability": "stable",
+        "runs": [
+          473.5,
+          493.5,
+          488.1,
+          510.3,
+          497.4,
+          487.5,
+          499.5,
+          485.3,
+          478.4,
+          480.3,
+          493.7,
+          476.1,
+          490.1,
+          479.3,
+          490.6,
+          489.5,
+          479.2,
+          488.8,
+          477.4,
+          473.5,
+          493.7,
+          488.2,
+          476.5,
+          495.9,
+          468.6,
+          496.5,
+          484.2,
+          489.8
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/sustained/result.json
index 4a35f301..9339f76d 100644
--- a/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_A_nvidia_vllm_47f5d58e_3f6269bb/sustained/result.json
@@ -422,7 +422,44 @@
       "sustained_throughput_tokens_per_sec": 486.6,
       "throttle_ratio": 0.918,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": 1.0
+      "ttft_p99_drift_ms": 1.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 486.6,
+        "std": 9.5,
+        "cv_pct": 1.95,
+        "stability": "stable",
+        "runs": [
+          473.5,
+          493.5,
+          488.1,
+          510.3,
+          497.4,
+          487.5,
+          499.5,
+          485.3,
+          478.4,
+          480.3,
+          493.7,
+          476.1,
+          490.1,
+          479.3,
+          490.6,
+          489.5,
+          479.2,
+          488.8,
+          477.4,
+          473.5,
+          493.7,
+          488.2,
+          476.5,
+          495.9,
+          468.6,
+          496.5,
+          484.2,
+          489.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/result.json
index 10356053..fedfa82a 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/result.json
@@ -381,7 +381,30 @@
       "sustained_throughput_tokens_per_sec": 484.8,
       "throttle_ratio": 0.887,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -202.4
+      "ttft_p99_drift_ms": -202.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 484.8,
+        "std": 15.2,
+        "cv_pct": 3.14,
+        "stability": "noisy",
+        "runs": [
+          443.9,
+          489.5,
+          495.1,
+          480.6,
+          500.6,
+          476.2,
+          492.7,
+          486.5,
+          482.4,
+          486.2,
+          499.3,
+          465.9,
+          499.2,
+          488.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/sustained/result.json
index d40598ee..99d5bd3a 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/bf16/sustained/result.json
@@ -272,7 +272,30 @@
       "sustained_throughput_tokens_per_sec": 484.8,
       "throttle_ratio": 0.887,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -202.4
+      "ttft_p99_drift_ms": -202.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 484.8,
+        "std": 15.2,
+        "cv_pct": 3.14,
+        "stability": "noisy",
+        "runs": [
+          443.9,
+          489.5,
+          495.1,
+          480.6,
+          500.6,
+          476.2,
+          492.7,
+          486.5,
+          482.4,
+          486.2,
+          499.3,
+          465.9,
+          499.2,
+          488.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/result.json
index 688d6ca8..a44887c8 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/result.json
@@ -381,7 +381,30 @@
       "sustained_throughput_tokens_per_sec": 494.9,
       "throttle_ratio": 0.92,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -111.0
+      "ttft_p99_drift_ms": -111.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 494.9,
+        "std": 12.6,
+        "cv_pct": 2.55,
+        "stability": "stable",
+        "runs": [
+          492.6,
+          482.1,
+          498.6,
+          498.3,
+          504.2,
+          472.5,
+          501.9,
+          505.4,
+          490.9,
+          483.1,
+          512.3,
+          477.3,
+          495.1,
+          513.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/sustained/result.json
index 21d4e7a4..489ee85f 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/fp8/sustained/result.json
@@ -272,7 +272,30 @@
       "sustained_throughput_tokens_per_sec": 494.9,
       "throttle_ratio": 0.92,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -111.0
+      "ttft_p99_drift_ms": -111.0,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 494.9,
+        "std": 12.6,
+        "cv_pct": 2.55,
+        "stability": "stable",
+        "runs": [
+          492.6,
+          482.1,
+          498.6,
+          498.3,
+          504.2,
+          472.5,
+          501.9,
+          505.4,
+          490.9,
+          483.1,
+          512.3,
+          477.3,
+          495.1,
+          513.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/result.json
index 7979d6b7..d0708288 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/result.json
@@ -381,7 +381,30 @@
       "sustained_throughput_tokens_per_sec": 648.8,
       "throttle_ratio": 0.934,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -206.6
+      "ttft_p99_drift_ms": -206.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 648.8,
+        "std": 9.5,
+        "cv_pct": 1.46,
+        "stability": "stable",
+        "runs": [
+          625.3,
+          669.7,
+          649.9,
+          651.4,
+          652.7,
+          645.9,
+          645.9,
+          642.4,
+          647.3,
+          647.3,
+          648.6,
+          648.7,
+          648.9,
+          658.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/sustained/result.json
index a01e5d58..f669eeb0 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w4a16/sustained/result.json
@@ -272,7 +272,30 @@
       "sustained_throughput_tokens_per_sec": 648.8,
       "throttle_ratio": 0.934,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -206.6
+      "ttft_p99_drift_ms": -206.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 648.8,
+        "std": 9.5,
+        "cv_pct": 1.46,
+        "stability": "stable",
+        "runs": [
+          625.3,
+          669.7,
+          649.9,
+          651.4,
+          652.7,
+          645.9,
+          645.9,
+          642.4,
+          647.3,
+          647.3,
+          648.6,
+          648.7,
+          648.9,
+          658.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/result.json
index 5aa2dc27..f00600e8 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/result.json
@@ -381,7 +381,30 @@
       "sustained_throughput_tokens_per_sec": 645.2,
       "throttle_ratio": 0.889,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -199.4
+      "ttft_p99_drift_ms": -199.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 645.2,
+        "std": 18.2,
+        "cv_pct": 2.82,
+        "stability": "stable",
+        "runs": [
+          593.7,
+          667.9,
+          642.7,
+          648.9,
+          646.6,
+          648.3,
+          657.5,
+          641.8,
+          648.5,
+          660.6,
+          643.0,
+          654.4,
+          623.1,
+          655.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/sustained/result.json
index cb0bb6a6..e8396426 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a16/sustained/result.json
@@ -272,7 +272,30 @@
       "sustained_throughput_tokens_per_sec": 645.2,
       "throttle_ratio": 0.889,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -199.4
+      "ttft_p99_drift_ms": -199.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 645.2,
+        "std": 18.2,
+        "cv_pct": 2.82,
+        "stability": "stable",
+        "runs": [
+          593.7,
+          667.9,
+          642.7,
+          648.9,
+          646.6,
+          648.3,
+          657.5,
+          641.8,
+          648.5,
+          660.6,
+          643.0,
+          654.4,
+          623.1,
+          655.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/result.json
index d2844329..c897b869 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/result.json
@@ -381,7 +381,30 @@
       "sustained_throughput_tokens_per_sec": 533.8,
       "throttle_ratio": 0.908,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -110.9
+      "ttft_p99_drift_ms": -110.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 533.8,
+        "std": 12.3,
+        "cv_pct": 2.3,
+        "stability": "stable",
+        "runs": [
+          501.4,
+          534.3,
+          552.1,
+          527.3,
+          528.0,
+          530.2,
+          539.2,
+          546.2,
+          526.4,
+          537.3,
+          526.7,
+          540.2,
+          542.8,
+          541.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/sustained/result.json
index b2534826..9627dbbe 100644
--- a/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_C_nvidia_vllm_47f5d58e_1bcdc710/w8a8/sustained/result.json
@@ -272,7 +272,30 @@
       "sustained_throughput_tokens_per_sec": 533.8,
       "throttle_ratio": 0.908,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -110.9
+      "ttft_p99_drift_ms": -110.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 533.8,
+        "std": 12.3,
+        "cv_pct": 2.3,
+        "stability": "stable",
+        "runs": [
+          501.4,
+          534.3,
+          552.1,
+          527.3,
+          528.0,
+          530.2,
+          539.2,
+          546.2,
+          526.4,
+          537.3,
+          526.7,
+          540.2,
+          542.8,
+          541.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/result.json b/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/result.json
index 8eb272c8..3b28d14f 100644
--- a/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/result.json
@@ -464,7 +464,44 @@
       "sustained_throughput_tokens_per_sec": 41.4,
       "throttle_ratio": 0.499,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 516.2
+      "ttft_p99_drift_ms": 516.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 41.4,
+        "std": 14.3,
+        "cv_pct": 34.47,
+        "stability": "high-variance",
+        "runs": [
+          34.1,
+          34.2,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.2,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          34.2,
+          68.3
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/sustained/result.json
index d6ad4f0d..839bccf7 100644
--- a/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_D_nvidia_vllm_47f5d58e_60c91bf0/sustained/result.json
@@ -422,7 +422,44 @@
       "sustained_throughput_tokens_per_sec": 41.4,
       "throttle_ratio": 0.499,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 516.2
+      "ttft_p99_drift_ms": 516.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 41.4,
+        "std": 14.3,
+        "cv_pct": 34.47,
+        "stability": "high-variance",
+        "runs": [
+          34.1,
+          34.2,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.2,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          34.1,
+          68.3,
+          34.1,
+          34.1,
+          34.1,
+          34.2,
+          68.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/result.json b/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/result.json
index 927da139..a3b8ad29 100644
--- a/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/result.json
@@ -359,7 +359,30 @@
       "sustained_throughput_tokens_per_sec": 1771.6,
       "throttle_ratio": 0.953,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -71.8
+      "ttft_p99_drift_ms": -71.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1771.6,
+        "std": 28.4,
+        "cv_pct": 1.6,
+        "stability": "stable",
+        "runs": [
+          1749.4,
+          1781.1,
+          1757.1,
+          1808.9,
+          1813.9,
+          1728.8,
+          1761.8,
+          1737.5,
+          1774.0,
+          1778.9,
+          1786.9,
+          1794.3,
+          1729.1,
+          1800.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/sustained/result.json b/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/sustained/result.json
index 1d501f43..7938869b 100644
--- a/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex1_suite_F_nvidia_vllm_47f5d58e_1e7ed8ca/sustained/result.json
@@ -272,7 +272,30 @@
       "sustained_throughput_tokens_per_sec": 1771.6,
       "throttle_ratio": 0.953,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -71.8
+      "ttft_p99_drift_ms": -71.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1771.6,
+        "std": 28.4,
+        "cv_pct": 1.6,
+        "stability": "stable",
+        "runs": [
+          1749.4,
+          1781.1,
+          1757.1,
+          1808.9,
+          1813.9,
+          1728.8,
+          1761.8,
+          1737.5,
+          1774.0,
+          1778.9,
+          1786.9,
+          1794.3,
+          1729.1,
+          1800.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/result.json b/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/result.json
index 29e70e1b..768a8fd0 100644
--- a/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/result.json
+++ b/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/result.json
@@ -564,7 +564,44 @@
       "sustained_throughput_tokens_per_sec": 176.0,
       "throttle_ratio": 0.847,
       "throttle_onset_minute": 5.0,
-      "ttft_p99_drift_ms": -9.3
+      "ttft_p99_drift_ms": -9.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 176.0,
+        "std": 6.5,
+        "cv_pct": 3.69,
+        "stability": "noisy",
+        "runs": [
+          174.4,
+          171.1,
+          181.0,
+          169.7,
+          177.7,
+          181.8,
+          176.6,
+          177.2,
+          177.7,
+          176.4,
+          177.1,
+          168.1,
+          175.3,
+          176.6,
+          179.5,
+          173.6,
+          185.1,
+          160.2,
+          186.4,
+          161.8,
+          189.2,
+          175.3,
+          176.9,
+          171.9,
+          179.6,
+          178.8,
+          180.0,
+          168.9
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 133.03,
diff --git a/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/sustained/result.json b/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/sustained/result.json
index 3061cd4f..1ac693dc 100644
--- a/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex8_suite_B_nvidia_vllm_47f5d58e_76ce4cd0/sustained/result.json
@@ -462,7 +462,44 @@
       "sustained_throughput_tokens_per_sec": 176.0,
       "throttle_ratio": 0.847,
       "throttle_onset_minute": 5.0,
-      "ttft_p99_drift_ms": -9.3
+      "ttft_p99_drift_ms": -9.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 176.0,
+        "std": 6.5,
+        "cv_pct": 3.69,
+        "stability": "noisy",
+        "runs": [
+          174.4,
+          171.1,
+          181.0,
+          169.7,
+          177.7,
+          181.8,
+          176.6,
+          177.2,
+          177.7,
+          176.4,
+          177.1,
+          168.1,
+          175.3,
+          176.6,
+          179.5,
+          173.6,
+          185.1,
+          160.2,
+          186.4,
+          161.8,
+          189.2,
+          175.3,
+          176.9,
+          171.9,
+          179.6,
+          178.8,
+          180.0,
+          168.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/result.json b/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/result.json
index e373079d..368cc57b 100644
--- a/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/result.json
+++ b/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/result.json
@@ -561,7 +561,44 @@
       "sustained_throughput_tokens_per_sec": 561.2,
       "throttle_ratio": 0.936,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -45.4
+      "ttft_p99_drift_ms": -45.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 561.2,
+        "std": 9.6,
+        "cv_pct": 1.71,
+        "stability": "stable",
+        "runs": [
+          561.4,
+          545.8,
+          565.7,
+          583.2,
+          554.6,
+          567.3,
+          556.8,
+          555.3,
+          571.4,
+          549.8,
+          561.4,
+          549.8,
+          565.6,
+          568.0,
+          563.2,
+          561.4,
+          562.6,
+          570.8,
+          546.6,
+          561.1,
+          576.4,
+          549.6,
+          550.5,
+          573.5,
+          556.8,
+          555.4,
+          555.4,
+          573.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/sustained/result.json b/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/sustained/result.json
index 73bdd2a1..29c90757 100644
--- a/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/sustained/result.json
+++ b/results/verified/nvidia_h20_3ex8_suite_G_nvidia_vllm_47f5d58e_7bd76bb5/sustained/result.json
@@ -462,7 +462,44 @@
       "sustained_throughput_tokens_per_sec": 561.2,
       "throttle_ratio": 0.936,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -45.4
+      "ttft_p99_drift_ms": -45.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 561.2,
+        "std": 9.6,
+        "cv_pct": 1.71,
+        "stability": "stable",
+        "runs": [
+          561.4,
+          545.8,
+          565.7,
+          583.2,
+          554.6,
+          567.3,
+          556.8,
+          555.3,
+          571.4,
+          549.8,
+          561.4,
+          549.8,
+          565.6,
+          568.0,
+          563.2,
+          561.4,
+          562.6,
+          570.8,
+          546.6,
+          561.1,
+          576.4,
+          549.6,
+          550.5,
+          573.5,
+          556.8,
+          555.4,
+          555.4,
+          573.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/result.json b/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/result.json
index 66048c9b..2d9a7a00 100644
--- a/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/result.json
+++ b/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/result.json
@@ -477,7 +477,44 @@
       "sustained_throughput_tokens_per_sec": 116.6,
       "throttle_ratio": 0.748,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 11.6
+      "ttft_p99_drift_ms": 11.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 116.6,
+        "std": 10.5,
+        "cv_pct": 8.99,
+        "stability": "high-variance",
+        "runs": [
+          98.7,
+          132.0,
+          105.3,
+          132.0,
+          117.6,
+          99.5,
+          128.9,
+          106.2,
+          126.7,
+          126.2,
+          108.1,
+          120.4,
+          105.7,
+          126.7,
+          117.5,
+          107.6,
+          120.4,
+          113.0,
+          126.8,
+          124.2,
+          103.4,
+          110.6,
+          121.4,
+          118.3,
+          125.2,
+          102.6,
+          110.5,
+          130.0
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/sustained/result.json b/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/sustained/result.json
index 49224331..49880a7c 100644
--- a/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/sustained/result.json
+++ b/results/verified/nvidia_l4x1_suite_A_nvidia_vllm_47f5d58e_b991b4c1/sustained/result.json
@@ -377,7 +377,44 @@
       "sustained_throughput_tokens_per_sec": 116.6,
       "throttle_ratio": 0.748,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 11.6
+      "ttft_p99_drift_ms": 11.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 116.6,
+        "std": 10.5,
+        "cv_pct": 8.99,
+        "stability": "high-variance",
+        "runs": [
+          98.7,
+          132.0,
+          105.3,
+          132.0,
+          117.6,
+          99.5,
+          128.9,
+          106.2,
+          126.7,
+          126.2,
+          108.1,
+          120.4,
+          105.7,
+          126.7,
+          117.5,
+          107.6,
+          120.4,
+          113.0,
+          126.8,
+          124.2,
+          103.4,
+          110.6,
+          121.4,
+          118.3,
+          125.2,
+          102.6,
+          110.5,
+          130.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/result.json b/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/result.json
index 7c987a20..b2d60369 100644
--- a/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/result.json
+++ b/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/result.json
@@ -314,7 +314,30 @@
       "sustained_throughput_tokens_per_sec": 2837.3,
       "throttle_ratio": 0.983,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -51.3
+      "ttft_p99_drift_ms": -51.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2837.3,
+        "std": 16.9,
+        "cv_pct": 0.6,
+        "stability": "stable",
+        "runs": [
+          2816.1,
+          2841.7,
+          2863.3,
+          2824.1,
+          2815.9,
+          2862.6,
+          2858.2,
+          2813.6,
+          2848.7,
+          2828.0,
+          2844.6,
+          2837.7,
+          2830.5,
+          2836.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/sustained/result.json b/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/sustained/result.json
index 5d1125ae..05e85ae7 100644
--- a/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/sustained/result.json
+++ b/results/verified/nvidia_l4x1_suite_F_nvidia_vllm_47f5d58e_d58fa923/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 2837.3,
       "throttle_ratio": 0.983,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -51.3
+      "ttft_p99_drift_ms": -51.3,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2837.3,
+        "std": 16.9,
+        "cv_pct": 0.6,
+        "stability": "stable",
+        "runs": [
+          2816.1,
+          2841.7,
+          2863.3,
+          2824.1,
+          2815.9,
+          2862.6,
+          2858.2,
+          2813.6,
+          2848.7,
+          2828.0,
+          2844.6,
+          2837.7,
+          2830.5,
+          2836.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/result.json b/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/result.json
index fcc6a4cd..a747a7d5 100644
--- a/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/result.json
+++ b/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/result.json
@@ -314,7 +314,30 @@
       "sustained_throughput_tokens_per_sec": 3880.8,
       "throttle_ratio": 0.982,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -48.2
+      "ttft_p99_drift_ms": -48.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 3880.8,
+        "std": 20.4,
+        "cv_pct": 0.52,
+        "stability": "stable",
+        "runs": [
+          3838.2,
+          3867.3,
+          3892.6,
+          3896.5,
+          3882.7,
+          3870.2,
+          3904.8,
+          3845.2,
+          3891.9,
+          3873.4,
+          3888.0,
+          3885.3,
+          3886.2,
+          3908.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/sustained/result.json b/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/sustained/result.json
index e983bcbf..67ac38d6 100644
--- a/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/sustained/result.json
+++ b/results/verified/nvidia_rtx_4000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_125c6b61/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 3880.8,
       "throttle_ratio": 0.982,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -48.2
+      "ttft_p99_drift_ms": -48.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 3880.8,
+        "std": 20.4,
+        "cv_pct": 0.52,
+        "stability": "stable",
+        "runs": [
+          3838.2,
+          3867.3,
+          3892.6,
+          3896.5,
+          3882.7,
+          3870.2,
+          3904.8,
+          3845.2,
+          3891.9,
+          3873.4,
+          3888.0,
+          3885.3,
+          3886.2,
+          3908.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/result.json
index 303f3772..9eee1e93 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/result.json
@@ -478,7 +478,44 @@
       "sustained_throughput_tokens_per_sec": 376.2,
       "throttle_ratio": 0.901,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": 18.3
+      "ttft_p99_drift_ms": 18.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 376.2,
+        "std": 8.1,
+        "cv_pct": 2.15,
+        "stability": "stable",
+        "runs": [
+          375.9,
+          376.2,
+          374.4,
+          394.4,
+          373.0,
+          365.6,
+          382.7,
+          363.9,
+          386.3,
+          375.6,
+          389.9,
+          355.4,
+          386.0,
+          372.6,
+          381.9,
+          376.1,
+          374.5,
+          370.4,
+          373.1,
+          380.1,
+          367.5,
+          384.4,
+          373.7,
+          376.4,
+          377.7,
+          376.4,
+          370.6,
+          379.9
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/sustained/result.json
index 4d90e875..ad67a5c4 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_A_nvidia_vllm_47f5d58e_bd3b5d27/sustained/result.json
@@ -377,7 +377,44 @@
       "sustained_throughput_tokens_per_sec": 376.2,
       "throttle_ratio": 0.901,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": 18.3
+      "ttft_p99_drift_ms": 18.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 376.2,
+        "std": 8.1,
+        "cv_pct": 2.15,
+        "stability": "stable",
+        "runs": [
+          375.9,
+          376.2,
+          374.4,
+          394.4,
+          373.0,
+          365.6,
+          382.7,
+          363.9,
+          386.3,
+          375.6,
+          389.9,
+          355.4,
+          386.0,
+          372.6,
+          381.9,
+          376.1,
+          374.5,
+          370.4,
+          373.1,
+          380.1,
+          367.5,
+          384.4,
+          373.7,
+          376.4,
+          377.7,
+          376.4,
+          370.6,
+          379.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/result.json
index bd6f82b1..f8c9b082 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/result.json
@@ -340,7 +340,30 @@
       "sustained_throughput_tokens_per_sec": 371.9,
       "throttle_ratio": 0.894,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -239.9
+      "ttft_p99_drift_ms": -239.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 371.9,
+        "std": 10.0,
+        "cv_pct": 2.69,
+        "stability": "stable",
+        "runs": [
+          348.3,
+          386.1,
+          365.8,
+          369.9,
+          372.0,
+          370.2,
+          372.6,
+          371.2,
+          389.7,
+          362.3,
+          380.4,
+          374.5,
+          373.2,
+          370.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/sustained/result.json
index b3717fde..001df410 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/bf16/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 371.9,
       "throttle_ratio": 0.894,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -239.9
+      "ttft_p99_drift_ms": -239.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 371.9,
+        "std": 10.0,
+        "cv_pct": 2.69,
+        "stability": "stable",
+        "runs": [
+          348.3,
+          386.1,
+          365.8,
+          369.9,
+          372.0,
+          370.2,
+          372.6,
+          371.2,
+          389.7,
+          362.3,
+          380.4,
+          374.5,
+          373.2,
+          370.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/result.json
index c557ea99..8b260dd4 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/result.json
@@ -340,7 +340,30 @@
       "sustained_throughput_tokens_per_sec": 586.1,
       "throttle_ratio": 0.969,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -131.6
+      "ttft_p99_drift_ms": -131.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 586.1,
+        "std": 5.3,
+        "cv_pct": 0.9,
+        "stability": "stable",
+        "runs": [
+          581.9,
+          582.3,
+          590.4,
+          582.1,
+          588.5,
+          592.1,
+          591.0,
+          588.4,
+          582.5,
+          573.7,
+          588.2,
+          582.6,
+          590.6,
+          590.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/sustained/result.json
index 3473e5c4..e2b9f0fb 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/fp8/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 586.1,
       "throttle_ratio": 0.969,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -131.6
+      "ttft_p99_drift_ms": -131.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 586.1,
+        "std": 5.3,
+        "cv_pct": 0.9,
+        "stability": "stable",
+        "runs": [
+          581.9,
+          582.3,
+          590.4,
+          582.1,
+          588.5,
+          592.1,
+          591.0,
+          588.4,
+          582.5,
+          573.7,
+          588.2,
+          582.6,
+          590.6,
+          590.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/result.json
index 10e8e5a7..4dc5165e 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/result.json
@@ -340,7 +340,30 @@
       "sustained_throughput_tokens_per_sec": 816.3,
       "throttle_ratio": 0.924,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -201.9
+      "ttft_p99_drift_ms": -201.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 816.3,
+        "std": 15.8,
+        "cv_pct": 1.94,
+        "stability": "stable",
+        "runs": [
+          778.0,
+          818.5,
+          841.8,
+          803.7,
+          816.3,
+          809.1,
+          829.1,
+          820.8,
+          806.2,
+          815.9,
+          826.8,
+          810.4,
+          813.1,
+          838.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/sustained/result.json
index 510302b5..4d15f3f6 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w4a16/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 816.3,
       "throttle_ratio": 0.924,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -201.9
+      "ttft_p99_drift_ms": -201.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 816.3,
+        "std": 15.8,
+        "cv_pct": 1.94,
+        "stability": "stable",
+        "runs": [
+          778.0,
+          818.5,
+          841.8,
+          803.7,
+          816.3,
+          809.1,
+          829.1,
+          820.8,
+          806.2,
+          815.9,
+          826.8,
+          810.4,
+          813.1,
+          838.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/result.json
index fafdfb44..a692cf2c 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/result.json
@@ -340,7 +340,30 @@
       "sustained_throughput_tokens_per_sec": 584.6,
       "throttle_ratio": 0.953,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -246.4
+      "ttft_p99_drift_ms": -246.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 584.6,
+        "std": 8.1,
+        "cv_pct": 1.38,
+        "stability": "stable",
+        "runs": [
+          565.5,
+          571.7,
+          589.7,
+          590.6,
+          581.9,
+          582.3,
+          590.4,
+          591.0,
+          590.5,
+          582.2,
+          582.2,
+          590.3,
+          582.1,
+          593.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/sustained/result.json
index 9616ccd0..b52408a5 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a16/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 584.6,
       "throttle_ratio": 0.953,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -246.4
+      "ttft_p99_drift_ms": -246.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 584.6,
+        "std": 8.1,
+        "cv_pct": 1.38,
+        "stability": "stable",
+        "runs": [
+          565.5,
+          571.7,
+          589.7,
+          590.6,
+          581.9,
+          582.3,
+          590.4,
+          591.0,
+          590.5,
+          582.2,
+          582.2,
+          590.3,
+          582.1,
+          593.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/result.json
index 454de8a9..a8762aec 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/result.json
@@ -340,7 +340,30 @@
       "sustained_throughput_tokens_per_sec": 534.5,
       "throttle_ratio": 0.952,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -110.6
+      "ttft_p99_drift_ms": -110.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 534.5,
+        "std": 5.8,
+        "cv_pct": 1.09,
+        "stability": "stable",
+        "runs": [
+          519.8,
+          531.9,
+          532.4,
+          535.4,
+          535.9,
+          536.8,
+          536.4,
+          537.7,
+          536.4,
+          533.3,
+          527.9,
+          537.8,
+          546.0,
+          534.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/sustained/result.json
index 357b1453..b7a08b4b 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_C_nvidia_vllm_47f5d58e_e60276e9/w8a8/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 534.5,
       "throttle_ratio": 0.952,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -110.6
+      "ttft_p99_drift_ms": -110.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 534.5,
+        "std": 5.8,
+        "cv_pct": 1.09,
+        "stability": "stable",
+        "runs": [
+          519.8,
+          531.9,
+          532.4,
+          535.4,
+          535.9,
+          536.8,
+          536.4,
+          537.7,
+          536.4,
+          533.3,
+          527.9,
+          537.8,
+          546.0,
+          534.9
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/result.json
index 7bb06c17..bab319b7 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/result.json
@@ -421,7 +421,44 @@
       "sustained_throughput_tokens_per_sec": 32.3,
       "throttle_ratio": 0.461,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 2424.7
+      "ttft_p99_drift_ms": 2424.7,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 32.3,
+        "std": 11.7,
+        "cv_pct": 36.18,
+        "stability": "high-variance",
+        "runs": [
+          25.6,
+          25.6,
+          25.6,
+          55.4,
+          25.6,
+          29.9,
+          25.6,
+          51.2,
+          25.6,
+          29.9,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6,
+          25.6,
+          55.5,
+          25.6,
+          29.9,
+          51.1,
+          25.7,
+          25.6,
+          29.9,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/sustained/result.json
index cb6e2f90..0667e527 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_D_nvidia_vllm_47f5d58e_42ab3af7/sustained/result.json
@@ -377,7 +377,44 @@
       "sustained_throughput_tokens_per_sec": 32.3,
       "throttle_ratio": 0.461,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 2424.7
+      "ttft_p99_drift_ms": 2424.7,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 32.3,
+        "std": 11.7,
+        "cv_pct": 36.18,
+        "stability": "high-variance",
+        "runs": [
+          25.6,
+          25.6,
+          25.6,
+          55.4,
+          25.6,
+          29.9,
+          25.6,
+          51.2,
+          25.6,
+          29.9,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6,
+          25.6,
+          55.5,
+          25.6,
+          29.9,
+          51.1,
+          25.7,
+          25.6,
+          29.9,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/result.json
index 502115f5..61f97879 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/result.json
@@ -314,7 +314,30 @@
       "sustained_throughput_tokens_per_sec": 2895.0,
       "throttle_ratio": 0.773,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -18.1
+      "ttft_p99_drift_ms": -18.1,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2895.0,
+        "std": 197.6,
+        "cv_pct": 6.83,
+        "stability": "noisy",
+        "runs": [
+          3526.2,
+          3053.6,
+          2843.1,
+          2823.8,
+          2814.7,
+          2834.9,
+          2805.7,
+          2814.8,
+          2790.3,
+          2824.4,
+          2862.5,
+          2960.7,
+          2848.3,
+          2726.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/sustained/result.json b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/sustained/result.json
index e5d7bf0c..97bc5729 100644
--- a/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/sustained/result.json
+++ b/results/verified/nvidia_rtx_6000_ada_generationx1_suite_F_nvidia_vllm_47f5d58e_2b905f5e/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 2895.0,
       "throttle_ratio": 0.773,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -18.1
+      "ttft_p99_drift_ms": -18.1,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2895.0,
+        "std": 197.6,
+        "cv_pct": 6.83,
+        "stability": "noisy",
+        "runs": [
+          3526.2,
+          3053.6,
+          2843.1,
+          2823.8,
+          2814.7,
+          2834.9,
+          2805.7,
+          2814.8,
+          2790.3,
+          2824.4,
+          2862.5,
+          2960.7,
+          2848.3,
+          2726.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/result.json b/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/result.json
index 726f575d..0cb9ba46 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/result.json
@@ -533,7 +533,44 @@
       "sustained_throughput_tokens_per_sec": 265.3,
       "throttle_ratio": 0.877,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": -18.4
+      "ttft_p99_drift_ms": -18.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 265.3,
+        "std": 8.6,
+        "cv_pct": 3.25,
+        "stability": "noisy",
+        "runs": [
+          259.8,
+          264.2,
+          272.7,
+          273.7,
+          262.9,
+          270.4,
+          247.2,
+          262.5,
+          273.3,
+          262.8,
+          272.8,
+          270.2,
+          252.1,
+          266.8,
+          263.8,
+          272.9,
+          258.3,
+          278.3,
+          270.2,
+          248.1,
+          262.5,
+          281.9,
+          254.3,
+          272.9,
+          270.5,
+          260.2,
+          259.2,
+          264.1
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/sustained/result.json
index b55019af..5c4fda4e 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_A_nvidia_vllm_47f5d58e_7cd0b745/sustained/result.json
@@ -432,7 +432,44 @@
       "sustained_throughput_tokens_per_sec": 265.3,
       "throttle_ratio": 0.877,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": -18.4
+      "ttft_p99_drift_ms": -18.4,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 265.3,
+        "std": 8.6,
+        "cv_pct": 3.25,
+        "stability": "noisy",
+        "runs": [
+          259.8,
+          264.2,
+          272.7,
+          273.7,
+          262.9,
+          270.4,
+          247.2,
+          262.5,
+          273.3,
+          262.8,
+          272.8,
+          270.2,
+          252.1,
+          266.8,
+          263.8,
+          272.9,
+          258.3,
+          278.3,
+          270.2,
+          248.1,
+          262.5,
+          281.9,
+          254.3,
+          272.9,
+          270.5,
+          260.2,
+          259.2,
+          264.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/result.json
index b108afd2..db10bf97 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/result.json
@@ -391,7 +391,30 @@
       "sustained_throughput_tokens_per_sec": 272.0,
       "throttle_ratio": 0.874,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -232.8
+      "ttft_p99_drift_ms": -232.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 272.0,
+        "std": 11.4,
+        "cv_pct": 4.19,
+        "stability": "noisy",
+        "runs": [
+          253.7,
+          275.2,
+          290.4,
+          285.6,
+          280.0,
+          282.0,
+          285.6,
+          271.1,
+          256.0,
+          265.8,
+          266.7,
+          262.5,
+          266.7,
+          266.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/sustained/result.json
index 06112e5a..01b6eee9 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/bf16/sustained/result.json
@@ -282,7 +282,30 @@
       "sustained_throughput_tokens_per_sec": 272.0,
       "throttle_ratio": 0.874,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -232.8
+      "ttft_p99_drift_ms": -232.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 272.0,
+        "std": 11.4,
+        "cv_pct": 4.19,
+        "stability": "noisy",
+        "runs": [
+          253.7,
+          275.2,
+          290.4,
+          285.6,
+          280.0,
+          282.0,
+          285.6,
+          271.1,
+          256.0,
+          265.8,
+          266.7,
+          262.5,
+          266.7,
+          266.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/result.json
index 9aa21947..083af211 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/result.json
@@ -391,7 +391,30 @@
       "sustained_throughput_tokens_per_sec": 435.5,
       "throttle_ratio": 0.891,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -477.8
+      "ttft_p99_drift_ms": -477.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 435.5,
+        "std": 11.5,
+        "cv_pct": 2.65,
+        "stability": "stable",
+        "runs": [
+          407.9,
+          438.6,
+          442.3,
+          439.1,
+          432.0,
+          441.9,
+          428.9,
+          438.9,
+          457.6,
+          425.9,
+          442.3,
+          430.0,
+          443.6,
+          427.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/sustained/result.json
index 65eabbca..8fc25e19 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/fp8/sustained/result.json
@@ -282,7 +282,30 @@
       "sustained_throughput_tokens_per_sec": 435.5,
       "throttle_ratio": 0.891,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -477.8
+      "ttft_p99_drift_ms": -477.8,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 435.5,
+        "std": 11.5,
+        "cv_pct": 2.65,
+        "stability": "stable",
+        "runs": [
+          407.9,
+          438.6,
+          442.3,
+          439.1,
+          432.0,
+          441.9,
+          428.9,
+          438.9,
+          457.6,
+          425.9,
+          442.3,
+          430.0,
+          443.6,
+          427.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/result.json
index 305ae8bf..70a58cb5 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/result.json
@@ -391,7 +391,30 @@
       "sustained_throughput_tokens_per_sec": 541.0,
       "throttle_ratio": 0.824,
       "throttle_onset_minute": 6.0,
-      "ttft_p99_drift_ms": -318.2
+      "ttft_p99_drift_ms": -318.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 541.0,
+        "std": 40.7,
+        "cv_pct": 7.53,
+        "stability": "noisy",
+        "runs": [
+          568.6,
+          600.0,
+          588.4,
+          608.5,
+          591.8,
+          535.1,
+          501.3,
+          513.5,
+          507.0,
+          515.9,
+          508.6,
+          521.6,
+          502.9,
+          510.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/sustained/result.json
index 0281b3d5..c9789133 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w4a16/sustained/result.json
@@ -282,7 +282,30 @@
       "sustained_throughput_tokens_per_sec": 541.0,
       "throttle_ratio": 0.824,
       "throttle_onset_minute": 6.0,
-      "ttft_p99_drift_ms": -318.2
+      "ttft_p99_drift_ms": -318.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 541.0,
+        "std": 40.7,
+        "cv_pct": 7.53,
+        "stability": "noisy",
+        "runs": [
+          568.6,
+          600.0,
+          588.4,
+          608.5,
+          591.8,
+          535.1,
+          501.3,
+          513.5,
+          507.0,
+          515.9,
+          508.6,
+          521.6,
+          502.9,
+          510.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/result.json
index 73ae3e59..958e98f2 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/result.json
@@ -391,7 +391,30 @@
       "sustained_throughput_tokens_per_sec": 433.8,
       "throttle_ratio": 0.91,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -471.4
+      "ttft_p99_drift_ms": -471.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 433.8,
+        "std": 9.6,
+        "cv_pct": 2.22,
+        "stability": "stable",
+        "runs": [
+          409.8,
+          439.3,
+          434.7,
+          433.3,
+          441.7,
+          440.2,
+          450.3,
+          424.3,
+          441.8,
+          433.8,
+          436.5,
+          429.7,
+          430.7,
+          427.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/sustained/result.json
index da8913c7..464d45b6 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a16/sustained/result.json
@@ -282,7 +282,30 @@
       "sustained_throughput_tokens_per_sec": 433.8,
       "throttle_ratio": 0.91,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -471.4
+      "ttft_p99_drift_ms": -471.4,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 433.8,
+        "std": 9.6,
+        "cv_pct": 2.22,
+        "stability": "stable",
+        "runs": [
+          409.8,
+          439.3,
+          434.7,
+          433.3,
+          441.7,
+          440.2,
+          450.3,
+          424.3,
+          441.8,
+          433.8,
+          436.5,
+          429.7,
+          430.7,
+          427.6
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/result.json
index e7a07246..ec510205 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/result.json
@@ -391,7 +391,30 @@
       "sustained_throughput_tokens_per_sec": 419.5,
       "throttle_ratio": 0.862,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -177.6
+      "ttft_p99_drift_ms": -177.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 419.5,
+        "std": 16.4,
+        "cv_pct": 3.9,
+        "stability": "noisy",
+        "runs": [
+          384.3,
+          426.5,
+          445.8,
+          403.7,
+          422.8,
+          435.2,
+          396.5,
+          437.6,
+          420.4,
+          418.3,
+          426.9,
+          417.6,
+          412.9,
+          424.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/sustained/result.json
index b6428506..e18178cf 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_C_nvidia_vllm_47f5d58e_b87c1621/w8a8/sustained/result.json
@@ -282,7 +282,30 @@
       "sustained_throughput_tokens_per_sec": 419.5,
       "throttle_ratio": 0.862,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -177.6
+      "ttft_p99_drift_ms": -177.6,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 419.5,
+        "std": 16.4,
+        "cv_pct": 3.9,
+        "stability": "noisy",
+        "runs": [
+          384.3,
+          426.5,
+          445.8,
+          403.7,
+          422.8,
+          435.2,
+          396.5,
+          437.6,
+          420.4,
+          418.3,
+          426.9,
+          417.6,
+          412.9,
+          424.8
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/result.json b/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/result.json
index f6f98fbe..70e12fab 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/result.json
@@ -476,7 +476,44 @@
       "sustained_throughput_tokens_per_sec": 30.5,
       "throttle_ratio": 0.461,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -34.2
+      "ttft_p99_drift_ms": -34.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 30.5,
+        "std": 10.5,
+        "cv_pct": 34.45,
+        "stability": "high-variance",
+        "runs": [
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.4,
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.5
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/sustained/result.json
index 42a7881e..1695366c 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_D_nvidia_vllm_47f5d58e_f2197473/sustained/result.json
@@ -432,7 +432,44 @@
       "sustained_throughput_tokens_per_sec": 30.5,
       "throttle_ratio": 0.461,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -34.2
+      "ttft_p99_drift_ms": -34.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 30.5,
+        "std": 10.5,
+        "cv_pct": 34.45,
+        "stability": "high-variance",
+        "runs": [
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.4,
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.5,
+          25.6,
+          25.6,
+          25.6,
+          29.9,
+          25.6,
+          25.6,
+          55.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/result.json b/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/result.json
index 7d98f82d..16faa2cf 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/result.json
@@ -369,7 +369,30 @@
       "sustained_throughput_tokens_per_sec": 1917.3,
       "throttle_ratio": 0.728,
       "throttle_onset_minute": 7.0,
-      "ttft_p99_drift_ms": -21.5
+      "ttft_p99_drift_ms": -21.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1917.3,
+        "std": 286.8,
+        "cv_pct": 14.96,
+        "stability": "high-variance",
+        "runs": [
+          2197.7,
+          2238.6,
+          2282.9,
+          2213.9,
+          2247.4,
+          2232.2,
+          1661.7,
+          1675.0,
+          1663.0,
+          1705.4,
+          1668.5,
+          1678.5,
+          1709.0,
+          1668.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/sustained/result.json b/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/sustained/result.json
index 9d66800d..96ccbaf5 100644
--- a/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x1_suite_F_nvidia_vllm_47f5d58e_a33d6eb3/sustained/result.json
@@ -282,7 +282,30 @@
       "sustained_throughput_tokens_per_sec": 1917.3,
       "throttle_ratio": 0.728,
       "throttle_onset_minute": 7.0,
-      "ttft_p99_drift_ms": -21.5
+      "ttft_p99_drift_ms": -21.5,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 1917.3,
+        "std": 286.8,
+        "cv_pct": 14.96,
+        "stability": "high-variance",
+        "runs": [
+          2197.7,
+          2238.6,
+          2282.9,
+          2213.9,
+          2247.4,
+          2232.2,
+          1661.7,
+          1675.0,
+          1663.0,
+          1705.4,
+          1668.5,
+          1678.5,
+          1709.0,
+          1668.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/result.json b/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/result.json
index a31b870e..5e4a0c73 100644
--- a/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/result.json
+++ b/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/result.json
@@ -574,7 +574,44 @@
       "sustained_throughput_tokens_per_sec": 105.3,
       "throttle_ratio": 0.764,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -30.2
+      "ttft_p99_drift_ms": -30.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 105.3,
+        "std": 6.8,
+        "cv_pct": 6.47,
+        "stability": "noisy",
+        "runs": [
+          108.7,
+          106.6,
+          102.3,
+          97.9,
+          111.4,
+          102.8,
+          101.8,
+          105.3,
+          113.3,
+          110.1,
+          100.9,
+          96.0,
+          108.5,
+          116.2,
+          88.8,
+          110.6,
+          111.3,
+          104.4,
+          107.8,
+          93.2,
+          109.5,
+          110.7,
+          101.6,
+          106.6,
+          103.9,
+          112.4,
+          93.3,
+          112.3
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 157.98,
diff --git a/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/sustained/result.json b/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/sustained/result.json
index d271ba19..6b364b39 100644
--- a/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x8_suite_B_nvidia_vllm_47f5d58e_0981ecf7/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 105.3,
       "throttle_ratio": 0.764,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -30.2
+      "ttft_p99_drift_ms": -30.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 105.3,
+        "std": 6.8,
+        "cv_pct": 6.47,
+        "stability": "noisy",
+        "runs": [
+          108.7,
+          106.6,
+          102.3,
+          97.9,
+          111.4,
+          102.8,
+          101.8,
+          105.3,
+          113.3,
+          110.1,
+          100.9,
+          96.0,
+          108.5,
+          116.2,
+          88.8,
+          110.6,
+          111.3,
+          104.4,
+          107.8,
+          93.2,
+          109.5,
+          110.7,
+          101.6,
+          106.6,
+          103.9,
+          112.4,
+          93.3,
+          112.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/result.json b/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/result.json
index ba4de6df..70f04083 100644
--- a/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/result.json
+++ b/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/result.json
@@ -571,7 +571,44 @@
       "sustained_throughput_tokens_per_sec": 343.0,
       "throttle_ratio": 0.893,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -70.2
+      "ttft_p99_drift_ms": -70.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 343.0,
+        "std": 10.9,
+        "cv_pct": 3.17,
+        "stability": "noisy",
+        "runs": [
+          341.7,
+          342.6,
+          326.1,
+          337.2,
+          353.4,
+          349.3,
+          330.5,
+          364.8,
+          341.4,
+          339.1,
+          352.7,
+          340.2,
+          328.8,
+          350.5,
+          342.9,
+          340.7,
+          332.5,
+          336.4,
+          357.0,
+          331.2,
+          364.2,
+          331.5,
+          349.5,
+          338.7,
+          350.9,
+          325.7,
+          356.4,
+          349.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/sustained/result.json b/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/sustained/result.json
index cd2c17af..53672ae3 100644
--- a/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/sustained/result.json
+++ b/results/verified/nvidia_rtx_a6000x8_suite_G_nvidia_vllm_47f5d58e_a8cf2a0f/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 343.0,
       "throttle_ratio": 0.893,
       "throttle_onset_minute": 4.0,
-      "ttft_p99_drift_ms": -70.2
+      "ttft_p99_drift_ms": -70.2,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 343.0,
+        "std": 10.9,
+        "cv_pct": 3.17,
+        "stability": "noisy",
+        "runs": [
+          341.7,
+          342.6,
+          326.1,
+          337.2,
+          353.4,
+          349.3,
+          330.5,
+          364.8,
+          341.4,
+          339.1,
+          352.7,
+          340.2,
+          328.8,
+          350.5,
+          342.9,
+          340.7,
+          332.5,
+          336.4,
+          357.0,
+          331.2,
+          364.2,
+          331.5,
+          349.5,
+          338.7,
+          350.9,
+          325.7,
+          356.4,
+          349.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/result.json b/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/result.json
index 06c3675f..9360106c 100644
--- a/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/result.json
+++ b/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/result.json
@@ -314,7 +314,30 @@
       "sustained_throughput_tokens_per_sec": 2006.9,
       "throttle_ratio": 0.982,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -156.7
+      "ttft_p99_drift_ms": -156.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2006.9,
+        "std": 10.1,
+        "cv_pct": 0.5,
+        "stability": "stable",
+        "runs": [
+          1999.4,
+          2004.0,
+          1998.1,
+          2014.0,
+          2013.6,
+          2003.3,
+          2015.3,
+          2006.1,
+          1997.0,
+          1997.9,
+          2013.0,
+          1999.4,
+          2033.6,
+          2002.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/sustained/result.json b/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/sustained/result.json
index 8f3b9484..9c1d9ecc 100644
--- a/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/sustained/result.json
+++ b/results/verified/tesla_t4x1_suite_F_nvidia_vllm_47f5d58e_4660bc0b/sustained/result.json
@@ -227,7 +227,30 @@
       "sustained_throughput_tokens_per_sec": 2006.9,
       "throttle_ratio": 0.982,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -156.7
+      "ttft_p99_drift_ms": -156.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2006.9,
+        "std": 10.1,
+        "cv_pct": 0.5,
+        "stability": "stable",
+        "runs": [
+          1999.4,
+          2004.0,
+          1998.1,
+          2014.0,
+          2013.6,
+          2003.3,
+          2015.3,
+          2006.1,
+          1997.0,
+          1997.9,
+          2013.0,
+          1999.4,
+          2033.6,
+          2002.5
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/result.json
index da09653a..769b621e 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/result.json
@@ -500,7 +500,44 @@
       "sustained_throughput_tokens_per_sec": 268.3,
       "throttle_ratio": 0.853,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -12.5
+      "ttft_p99_drift_ms": -12.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 268.3,
+        "std": 11.1,
+        "cv_pct": 4.12,
+        "stability": "noisy",
+        "runs": [
+          265.0,
+          294.4,
+          276.4,
+          269.7,
+          295.7,
+          281.2,
+          281.4,
+          271.3,
+          263.5,
+          265.3,
+          255.6,
+          262.4,
+          276.5,
+          255.3,
+          263.9,
+          254.9,
+          282.2,
+          262.4,
+          271.5,
+          252.3,
+          262.6,
+          268.9,
+          261.8,
+          269.2,
+          260.4,
+          262.4,
+          269.1,
+          257.7
+        ]
+      }
     },
     "speculative": {
       "results_by_concurrency": [
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/sustained/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/sustained/result.json
index adbaa643..2060ee84 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/sustained/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_A_nvidia_vllm_47f5d58e_48261ecc/sustained/result.json
@@ -402,7 +402,44 @@
       "sustained_throughput_tokens_per_sec": 268.3,
       "throttle_ratio": 0.853,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": -12.5
+      "ttft_p99_drift_ms": -12.5,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 268.3,
+        "std": 11.1,
+        "cv_pct": 4.12,
+        "stability": "noisy",
+        "runs": [
+          265.0,
+          294.4,
+          276.4,
+          269.7,
+          295.7,
+          281.2,
+          281.4,
+          271.3,
+          263.5,
+          265.3,
+          255.6,
+          262.4,
+          276.5,
+          255.3,
+          263.9,
+          254.9,
+          282.2,
+          262.4,
+          271.5,
+          252.3,
+          262.6,
+          268.9,
+          261.8,
+          269.2,
+          260.4,
+          262.4,
+          269.1,
+          257.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/result.json
index b1111115..b0451f0e 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 265.9,
       "throttle_ratio": 0.864,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -315.2
+      "ttft_p99_drift_ms": -315.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 265.9,
+        "std": 11.7,
+        "cv_pct": 4.41,
+        "stability": "noisy",
+        "runs": [
+          253.9,
+          268.8,
+          273.6,
+          288.3,
+          272.3,
+          261.3,
+          277.1,
+          252.4,
+          268.1,
+          249.0,
+          278.5,
+          253.5,
+          269.7,
+          256.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/sustained/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/sustained/result.json
index dd11245f..22e3245f 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/sustained/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/fp16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 265.9,
       "throttle_ratio": 0.864,
       "throttle_onset_minute": 1.0,
-      "ttft_p99_drift_ms": -315.2
+      "ttft_p99_drift_ms": -315.2,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 265.9,
+        "std": 11.7,
+        "cv_pct": 4.41,
+        "stability": "noisy",
+        "runs": [
+          253.9,
+          268.8,
+          273.6,
+          288.3,
+          272.3,
+          261.3,
+          277.1,
+          252.4,
+          268.1,
+          249.0,
+          278.5,
+          253.5,
+          269.7,
+          256.7
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/result.json
index 7fd24ec5..b93e3207 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/result.json
@@ -361,7 +361,30 @@
       "sustained_throughput_tokens_per_sec": 416.4,
       "throttle_ratio": 0.915,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -335.7
+      "ttft_p99_drift_ms": -335.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 416.4,
+        "std": 10.9,
+        "cv_pct": 2.62,
+        "stability": "stable",
+        "runs": [
+          400.3,
+          404.0,
+          414.5,
+          428.4,
+          407.3,
+          435.2,
+          418.8,
+          418.2,
+          419.1,
+          413.0,
+          426.8,
+          422.6,
+          423.5,
+          398.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/sustained/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/sustained/result.json
index 003f7bcc..8f855f82 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/sustained/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_C_nvidia_vllm_47f5d58e_b957e789/w4a16/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 416.4,
       "throttle_ratio": 0.915,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -335.7
+      "ttft_p99_drift_ms": -335.7,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 416.4,
+        "std": 10.9,
+        "cv_pct": 2.62,
+        "stability": "stable",
+        "runs": [
+          400.3,
+          404.0,
+          414.5,
+          428.4,
+          407.3,
+          435.2,
+          418.8,
+          418.2,
+          419.1,
+          413.0,
+          426.8,
+          422.6,
+          423.5,
+          398.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/result.json
index 17b35079..5e3dda03 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/result.json
@@ -444,7 +444,44 @@
       "sustained_throughput_tokens_per_sec": 14.9,
       "throttle_ratio": 0.399,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 34216.0
+      "ttft_p99_drift_ms": 34216.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 14.9,
+        "std": 4.4,
+        "cv_pct": 29.72,
+        "stability": "high-variance",
+        "runs": [
+          17.1,
+          17.1,
+          8.5,
+          17.1,
+          17.1,
+          8.5,
+          17.1,
+          21.3,
+          8.5,
+          17.1,
+          17.1,
+          8.5,
+          21.3,
+          17.1,
+          8.5,
+          17.1,
+          17.1,
+          17.1,
+          8.5,
+          17.1,
+          17.1,
+          8.5,
+          21.3,
+          17.1,
+          8.5,
+          17.1,
+          12.8,
+          17.1
+        ]
+      }
     },
     "online": {
       "sla_ttft_ms": 5000,
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/sustained/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/sustained/result.json
index 07f35034..7862c4c8 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/sustained/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_D_nvidia_vllm_47f5d58e_6eb549a8/sustained/result.json
@@ -402,7 +402,44 @@
       "sustained_throughput_tokens_per_sec": 14.9,
       "throttle_ratio": 0.399,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 34216.0
+      "ttft_p99_drift_ms": 34216.0,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 14.9,
+        "std": 4.4,
+        "cv_pct": 29.72,
+        "stability": "high-variance",
+        "runs": [
+          17.1,
+          17.1,
+          8.5,
+          17.1,
+          17.1,
+          8.5,
+          17.1,
+          21.3,
+          8.5,
+          17.1,
+          17.1,
+          8.5,
+          21.3,
+          17.1,
+          8.5,
+          17.1,
+          17.1,
+          17.1,
+          8.5,
+          17.1,
+          17.1,
+          8.5,
+          21.3,
+          17.1,
+          8.5,
+          17.1,
+          12.8,
+          17.1
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/result.json
index 890bb38f..08b68576 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/result.json
@@ -336,7 +336,30 @@
       "sustained_throughput_tokens_per_sec": 2789.7,
       "throttle_ratio": 0.927,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -83.9
+      "ttft_p99_drift_ms": -83.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2789.7,
+        "std": 50.6,
+        "cv_pct": 1.81,
+        "stability": "stable",
+        "runs": [
+          2797.6,
+          2649.7,
+          2789.6,
+          2783.6,
+          2742.5,
+          2781.3,
+          2791.3,
+          2763.1,
+          2817.4,
+          2817.8,
+          2842.1,
+          2857.4,
+          2830.4,
+          2792.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/sustained/result.json b/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/sustained/result.json
index 2091a26b..39e3777b 100644
--- a/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/sustained/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx1_suite_F_nvidia_vllm_47f5d58e_04fce6f6/sustained/result.json
@@ -252,7 +252,30 @@
       "sustained_throughput_tokens_per_sec": 2789.7,
       "throttle_ratio": 0.927,
       "throttle_onset_minute": null,
-      "ttft_p99_drift_ms": -83.9
+      "ttft_p99_drift_ms": -83.9,
+      "throughput_post_warmup_reliability": {
+        "n": 14,
+        "mean": 2789.7,
+        "std": 50.6,
+        "cv_pct": 1.81,
+        "stability": "stable",
+        "runs": [
+          2797.6,
+          2649.7,
+          2789.6,
+          2783.6,
+          2742.5,
+          2781.3,
+          2791.3,
+          2763.1,
+          2817.4,
+          2817.8,
+          2842.1,
+          2857.4,
+          2830.4,
+          2792.3
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/result.json b/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/result.json
index 6536a548..27dd8c63 100644
--- a/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/result.json
@@ -571,7 +571,44 @@
       "sustained_throughput_tokens_per_sec": 92.8,
       "throttle_ratio": 0.749,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 34.6
+      "ttft_p99_drift_ms": 34.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 92.8,
+        "std": 8.0,
+        "cv_pct": 8.63,
+        "stability": "high-variance",
+        "runs": [
+          86.6,
+          95.9,
+          92.6,
+          103.3,
+          80.9,
+          93.4,
+          92.7,
+          84.7,
+          102.0,
+          89.7,
+          88.6,
+          94.3,
+          92.8,
+          106.5,
+          81.9,
+          92.7,
+          100.4,
+          83.5,
+          99.3,
+          86.1,
+          97.7,
+          82.1,
+          97.0,
+          102.8,
+          79.8,
+          106.2,
+          84.7,
+          99.2
+        ]
+      }
     },
     "interactive": {
       "ttft_ms_p50": 225.66,
diff --git a/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/sustained/result.json b/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/sustained/result.json
index 23d04c69..e34441a4 100644
--- a/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/sustained/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx8_suite_B_nvidia_vllm_47f5d58e_48f19c22/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 92.8,
       "throttle_ratio": 0.749,
       "throttle_onset_minute": 2.0,
-      "ttft_p99_drift_ms": 34.6
+      "ttft_p99_drift_ms": 34.6,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 92.8,
+        "std": 8.0,
+        "cv_pct": 8.63,
+        "stability": "high-variance",
+        "runs": [
+          86.6,
+          95.9,
+          92.6,
+          103.3,
+          80.9,
+          93.4,
+          92.7,
+          84.7,
+          102.0,
+          89.7,
+          88.6,
+          94.3,
+          92.8,
+          106.5,
+          81.9,
+          92.7,
+          100.4,
+          83.5,
+          99.3,
+          86.1,
+          97.7,
+          82.1,
+          97.0,
+          102.8,
+          79.8,
+          106.2,
+          84.7,
+          99.2
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/result.json b/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/result.json
index 4c93f34e..3d656bed 100644
--- a/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/result.json
@@ -571,7 +571,44 @@
       "sustained_throughput_tokens_per_sec": 293.9,
       "throttle_ratio": 0.811,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": 22.3
+      "ttft_p99_drift_ms": 22.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 293.9,
+        "std": 15.7,
+        "cv_pct": 5.35,
+        "stability": "noisy",
+        "runs": [
+          311.7,
+          301.1,
+          301.5,
+          293.5,
+          300.5,
+          320.6,
+          285.0,
+          308.8,
+          297.6,
+          290.8,
+          304.3,
+          316.8,
+          293.8,
+          287.4,
+          316.6,
+          305.1,
+          296.0,
+          298.7,
+          260.1,
+          298.2,
+          284.7,
+          274.4,
+          293.2,
+          267.2,
+          271.3,
+          297.0,
+          262.9,
+          291.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/sustained/result.json b/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/sustained/result.json
index 1804e6c1..ab2680be 100644
--- a/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/sustained/result.json
+++ b/results/verified/tesla_v100s_pcie_32gbx8_suite_G_nvidia_vllm_47f5d58e_2ef567be/sustained/result.json
@@ -472,7 +472,44 @@
       "sustained_throughput_tokens_per_sec": 293.9,
       "throttle_ratio": 0.811,
       "throttle_onset_minute": 8.0,
-      "ttft_p99_drift_ms": 22.3
+      "ttft_p99_drift_ms": 22.3,
+      "throughput_post_warmup_reliability": {
+        "n": 28,
+        "mean": 293.9,
+        "std": 15.7,
+        "cv_pct": 5.35,
+        "stability": "noisy",
+        "runs": [
+          311.7,
+          301.1,
+          301.5,
+          293.5,
+          300.5,
+          320.6,
+          285.0,
+          308.8,
+          297.6,
+          290.8,
+          304.3,
+          316.8,
+          293.8,
+          287.4,
+          316.6,
+          305.1,
+          296.0,
+          298.7,
+          260.1,
+          298.2,
+          284.7,
+          274.4,
+          293.2,
+          267.2,
+          271.3,
+          297.0,
+          262.9,
+          291.0
+        ]
+      }
     }
   },
   "accuracy": {
diff --git a/schema/env.schema.json b/schema/env.schema.json
index e80cd942..4d90a791 100644
--- a/schema/env.schema.json
+++ b/schema/env.schema.json
@@ -73,6 +73,11 @@
     "intra_node_interconnect": {
       "type": ["string","null"],
       "description": "Intra-node GPU interconnect detected, e.g. 'NVLink', or null if not detected"
+    },
+    "vendor_details": {
+      "type": ["object","null"],
+      "additionalProperties": true,
+      "description": "Optional bag of vendor-specific environment fields that don't fit any unified schema (e.g. NVIDIA NVML clocks, AMD ROCm-SMI counters, Ascend HCCL link health, Apple Metal version). The schema is deliberately permissive — each vendor's platforms/<vendor>.py decides what to record. The leaderboard UI renders this dict as a flat key→value list, omitting null/empty values, and never tries to unify across vendors."
     }
   }
 }
diff --git a/schema/suite.schema.json b/schema/suite.schema.json
index 1367fe0d..7309b58a 100644
--- a/schema/suite.schema.json
+++ b/schema/suite.schema.json
@@ -117,11 +117,23 @@
     "online_sla_ttft_ms":         { "type": ["integer", "null"], "minimum": 1 },
     "online_sla_ttft_ms_relaxed": { "type": ["integer", "null"], "minimum": 1 },
     "online_request_count":       { "type": ["integer", "null"], "minimum": 1 },
-    "online_warmup_runs":         { "type": "integer", "minimum": 0 },
+    "online_warmup_runs":         {
+      "type": "integer", "minimum": 0,
+      "description": "DEPRECATED. Previously unused — kept only to silence schema warnings on older suites. Use online_warmup_requests instead."
+    },
+    "online_warmup_requests":     {
+      "type": "integer", "minimum": 0,
+      "description": "Number of dummy requests fired sequentially before the online QPS sweep. Results are discarded. Used to JIT-compile kernels and prime the engine on cold start. Defaults to 10 if not set."
+    },
 
     "interactive_request_count":  { "type": ["integer", "null"], "minimum": 1 },
     "interactive_warmup_runs":    { "type": "integer", "minimum": 0 },
 
+    "burst_warmup_requests":      {
+      "type": "integer", "minimum": 0,
+      "description": "Number of dummy requests fired sequentially before the first burst cycle. Results are discarded. Defaults to 10 if not set."
+    },
+
     "sustained_concurrency":      { "type": "integer", "minimum": 1 },
     "duration_minutes":           { "type": "number",  "minimum": 0 },
     "sample_interval_seconds":    { "type": "number",  "minimum": 0 },
diff --git a/suites/suite_A/suite.json b/suites/suite_A/suite.json
index d8f6914a..e902022a 100644
--- a/suites/suite_A/suite.json
+++ b/suites/suite_A/suite.json
@@ -29,8 +29,9 @@
   "num_runs": 3,
   "warmup_runs": 1,
   "warmup_minutes": 2,
-  "online_warmup_runs": 0,
-  "interactive_warmup_runs": 0 ,
+  "online_warmup_requests": 10,
+  "burst_warmup_requests": 10,
+  "interactive_warmup_runs": 0,
   "accuracy_threshold_delta": 0.1,
   "request_count": 100,
   "_request_count_note": "offline uses request_count (100), online uses online_request_count (300, minimum for robust p99), interactive uses interactive_request_count (150, minimum for robust p95)",
diff --git a/suites/suite_B/suite.json b/suites/suite_B/suite.json
index b331dbd8..87ba98f0 100644
--- a/suites/suite_B/suite.json
+++ b/suites/suite_B/suite.json
@@ -31,7 +31,8 @@
   "num_runs": 3,
   "warmup_runs": 1,
   "warmup_minutes": 2,
-  "online_warmup_runs": 0,
+  "online_warmup_requests": 10,
+  "burst_warmup_requests": 10,
   "interactive_warmup_runs": 0,
   "accuracy_threshold_delta": 0.1,
   "request_count": 100,
diff --git a/suites/suite_D/suite.json b/suites/suite_D/suite.json
index 0f93adc9..2f439d0a 100644
--- a/suites/suite_D/suite.json
+++ b/suites/suite_D/suite.json
@@ -29,7 +29,7 @@
   "num_runs": 2,
   "warmup_runs": 1,
   "warmup_minutes": 2,
-  "online_warmup_runs": 0,
+  "online_warmup_requests": 10,
   "interactive_warmup_runs": 0,
   "accuracy_threshold_delta": 0.1,
   "request_count": 50,
diff --git a/suites/suite_E/suite.json b/suites/suite_E/suite.json
index dfaab1bb..6a3de48d 100644
--- a/suites/suite_E/suite.json
+++ b/suites/suite_E/suite.json
@@ -32,7 +32,7 @@
   "online_sla_ttft_ms_relaxed": null,
   "num_runs": 3,
   "warmup_runs": 1,
-  "online_warmup_runs": 0,
+  "online_warmup_requests": 10,
   "interactive_warmup_runs": 0,
   "accuracy_threshold_delta": 0.1,
   "request_count": 150,
diff --git a/suites/suite_F/suite.json b/suites/suite_F/suite.json
index 17f3a9b4..67851df0 100644
--- a/suites/suite_F/suite.json
+++ b/suites/suite_F/suite.json
@@ -33,7 +33,7 @@
   "num_runs": 3,
   "warmup_runs": 1,
   "warmup_minutes": 1,
-  "online_warmup_runs": 0,
+  "online_warmup_requests": 10,
   "interactive_warmup_runs": 0,
   "accuracy_threshold_delta": 0.1,
   "_accuracy_note": "Qwen2.5-0.5B scores ~0.35–0.40 on MMLU by design (small model). The threshold detects broken quantization or misconfigured precision, not model quality.",
diff --git a/suites/suite_G/suite.json b/suites/suite_G/suite.json
index 7ef5aae2..0a339dab 100644
--- a/suites/suite_G/suite.json
+++ b/suites/suite_G/suite.json
@@ -31,7 +31,7 @@
   "num_runs": 3,
   "warmup_runs": 1,
   "warmup_minutes": 2,
-  "online_warmup_runs": 0,
+  "online_warmup_requests": 10,
   "interactive_warmup_runs": 0,
   "accuracy_threshold_delta": 0.1,
   "request_count": 100,