Execution Board: NL paper-alignment milestone tracker

[georgepullen]

NL Paper-Alignment Execution Board

This is the execution tracker for milestone NL Paper-Alignment Execution Board.
Project board: https://github.com/users/georgepullen/projects/1

Mandatory Protocol (applies to every linked ticket)

Before coding, the assigned agent must post a first comment proving they read:

• reports/NL_IMPLEMENTATION_ORACLE.md
• reports/paper/NL-print.extracted.clean.txt
• Ticket-specific required docs/code listed in that ticket

The first comment must include:

1. Oracle summary relevant to the ticket.
2. Paper equation mapping relevant to the ticket.
3. Exact files they will modify.
4. Risks and non-goals.

No implementation work should begin before this comment.

Active Queue

Phase 0

• Phase 0: Establish canonical paper-faithful baselines and lock reference metrics #1
• Phase 0: Add fidelity telemetry validator and metadata gates #2

Phase 1 (optimizer equation fidelity)

• Phase 1: Implement Eq. 21-22 delta-rule momentum variant #3
• Phase 1: Implement Eq. 23 DMGD with MLP-parameterized momentum #4
• Phase 1: Implement Eq. 24-style Newton-Schulz inner momentum variant (muon_ns) #5

Phase 2 (CMS + regression gates)

• Phase 2: Implement Eq. 28-29 CMS projection optimizer path #6
• Phase 2: Add equation-level fidelity regression suite for optimizer and CMS updates #7

Phase 3 (training semantics)

• Phase 3: Implement differentiable online-write mode (boundary-state gradients) #8

Phase 4 (distributed parity)

• Phase 4: Port paper-faithful online/per-layer path to FSDP with parity tests #9

Phase 5 (RunPod execution + scale)

• RunPod: Add reproducible pod setup, storage, and checkpoint-resume playbook #10
• Phase 5: Run RunPod pilot ablation matrix (5k/25k) for paper-fidelity paths #11
• Phase 5: Execute mid/target faithful runs and publish final alignment report #12

Board Rules

• One major phase active at a time unless blocked by infra.
• Every ticket closure must include:
  • exact commands run
  • checkpoint paths
  • eval JSON paths
  • acceptance criteria evidence
• Any deviation from oracle/paper must be explicitly documented in ticket comments.

[georgepullen]

Execution update (in-order progress):

• Phase 0: Establish canonical paper-faithful baselines and lock reference metrics #1 Phase-0 baseline orchestration + packaging scaffold landed (commands, eval wiring, bundle manifest/report generation).
• Phase 0: Add fidelity telemetry validator and metadata gates #2 telemetry quality-gate validator landed (script + fixtures + CI job + training surprise telemetry keys).
• RunPod: Add reproducible pod setup, storage, and checkpoint-resume playbook #10 RunPod reproducibility playbook landed (docs + bootstrap/sync/validate/resume helper scripts).
• Phase 1: Implement Eq. 21-22 delta-rule momentum variant #3 started: delta-rule inner optimizer variant implemented with tests + ablation config.

Current lane state:

• In Progress: Phase 0: Establish canonical paper-faithful baselines and lock reference metrics #1, Phase 0: Add fidelity telemetry validator and metadata gates #2, Phase 1: Implement Eq. 21-22 delta-rule momentum variant #3, RunPod: Add reproducible pod setup, storage, and checkpoint-resume playbook #10, Execution Board: NL paper-alignment milestone tracker #13
• Todo: Phase 1: Implement Eq. 23 DMGD with MLP-parameterized momentum #4-Phase 4: Port paper-faithful online/per-layer path to FSDP with parity tests #9, Phase 5: Run RunPod pilot ablation matrix (5k/25k) for paper-fidelity paths #11-Phase 5: Execute mid/target faithful runs and publish final alignment report #12

Pending close criteria requiring runtime execution evidence:

• Full 2k Phase-0 baseline runs and packaged artifacts.
• RunPod fresh-pod proof and resume drill evidence from target environment.
• 5k pilot finite run for delta-rule variant.