v5.2.6 — Accuracy + Consistency Fixes

v5.2.6 — Accuracy + Consistency Fixes

This tag bundles two reviews against v5.2.5: the original 5 accuracy fixes (editorial review) and 6 rendering/consistency fixes (#2). Every fix is verified against the protocol document's own definitions — no speculation.

---

Part 1 — Accuracy Patch (editorial review)

Five independently verifiable accuracy fixes found during systematic editorial review of v5.2.5.

Changes

ai-verification-protocol.md

1. §3.6 — six sub-signals (m, o, b, f, s, t) → seven sub-signals (m, o, b, f, s, t, d)
2. §8.2 — Same fix. Signal d (doc coverage) was added in v5.0 (§2.9, weight 0.05). Seven signals are defined in §3.1, but two sections still referenced "six."

index.html
3. Carousel slide 1 — η 0.94→0.97, ρ 0.04→0.03. Per §3.3, η ≥ 0.95 is required for AutoApprove. η=0.94 falls in the 0.80–0.95 band → HumanReviewRecommended. Cert now matches its verdict.
4. Carousel slide 4 — Removed false "PR size 1,892 LOC exceeds 1,500 hard cap". Per §0.3, the hard cap is 5,000 LOC; 1,892 falls in 1,501–5,000 (capped at HumanReviewRecommended). The actual CannotVerify trigger is ρ > 0.30.
5. Body text — $0.0015 → $0.015 per-PR generation cost. $50 ÷ $0.0015 = 33,333:1, not the claimed ~3,300:1. Fixed to $0.015 (50/0.015 = 3,333 ≈ 3,300).

Verification

Fix	Source of Truth	Result
Signal count	§3.1 defines 7 signals	7 ✓
η=0.97 verdict	§3.3 η≥0.95 → AutoApprove	Correct ✓
Size cap removed	§0.3 hard cap at 5,000	Removed ✓
Math: $0.015	50/0.015=3,333≈3,300	Matches ✓

---

Part 2 — Rendering & Consistency Fixes (PR #2)

Six surgical fixes from a two-pass reassessment. No protocol logic changed — these restore document self-consistency and fix markdown rendering. Finding 4 was redirected during review (the original target, Appendix A, was already correct; the real defect was the §3.5 table).

Changes

ai-verification-protocol.md

1. §7.5 — Moved the mandatory same-family constraint out of the code fence; it was rendering as gray monospace with literal **asterisks**, now reads as body prose.
2. §6.1, §6.6 — Repaired three table rows missing their leading | (Adversarial Surface, Documentation Coverage, Documentation generation) — they were misrendering.
3. §6.7 / §7.5 — Scoped the 3-attempt repair cap explicitly as per-PR (shared budget across all auto-repairable findings), not per-finding.
4. §3.5 — Added the missing spec_independence row to the ρ sub-signal table. The four prior rows summed to 0.25 while the stated cap is 0.30 — the fifth +0.05 row closes the gap. (Appendix A, the schema, and §6.5.1 already listed it; the §3.5 table was the lone omission.)
5. §3.5 — Added a back-reference to the §0.2 unknown-generator_identity fallback (family/version → maximum contribution).
6. §3.8 — Annotated the pr_size_cap ceiling as HumanReviewRequired per §0.3 inside the max_severity() block.

index.html / og.svg — Reconciled the two version strings the accuracy patch missed (5.2.5 → 5.2.6 in the CTA footer and the social card), so the version is now consistent across every carrier.

Verification

Finding	Source of Truth	Result
§7.5 prose outside fence	markdown fence balance	✓
3 table rows repaired	leading `	` present
Cap scoped per-PR	§6.7 + §7.5 agree	✓
§3.5 table = 5 rows	contributions sum to 0.30 cap	✓
§3.5 → §0.2 cross-ref	present	✓
§3.8 ceiling annotated	HumanReviewRequired per §0.3	✓
Version consistency	zero residual 5.2.5	✓

Protocol self-run

PR #2 was verified by running the protocol on itself (five-agent pipeline). Verdict: HumanReviewRequired — content clean (η_raw 0.93, 8/8 oracles), but ρ=0.20 from single-family self-verification (Claude verifying Claude, no provenance) pulled η to 0.73. The protocol correctly distrusted its own monoculture pipeline. Full certificate.

---

Full diff: v5.2.5...v5.2.6

v5.2.5 — The AI Verification Protocol

🚀 The AI Verification Protocol

Diagnose, repair, and measure — the operational answer to AI verification debt.

A multi-agent pipeline specification and system prompt that quantifies verification debt, derives η from observable signals, tracks Ci/Cv ratios, and orchestrates a five-agent review pipeline with provenance attestation.

📋 What It Does

Feature	Detail
9 verification axes	Semantic, behavioral, security, structural, fuzzing, dependency, provenance, adversarial, documentation
5 pipeline agents	A (generator) → B (reviewer) → C (contract) → D (fuzzer) → E (certificate)
η from signals	Mechanical efficiency score from 7 observable signals (m, o, b, f, s, t, d)
ρ correlation penalty	Quantifies verifier-generator dependency; ρ > 0.30 → CannotVerify
Ci/Cv ratio	Cost-to-Verify ÷ Cost-to-Implement per PR — the metric that matters
Active Repair Mode	Auto-generates tests, docs, and type fixes with 5-gate verification
Machine-readable certificates	JSON + in-toto attestation; markdown rendering for humans
Meta-audit loop	5% monthly sampling, Brier calibration, weight auto-recalibration

🌐 Landing Page

Live at vprotocol.21no.de — includes:

• Certificate carousel showcasing all 4 verdict outcomes (AutoApprove → CannotVerify)
• Five Whys root-cause analysis with whitepaper citations
• Pipeline flow and feature cards

📖 Reads

• Protocol: ai-verification-protocol.md — 14 sections, 2 appendices, 1,167 lines
• Companion whitepaper: The AI Verification Debt

📦 Contents

index.html · ai-verification-protocol.md · README.md · CNAME · og.svg · LICENSE