Fair clock — public rules (NoCap)
Audience: strangers auditing speed claims.
Evidence state: 2026-07-16.
Companion glossary: GLOSSARY.html.
Quality-gate primer: disjoint-val-explained.html.
Private ops protocol (agents): nocap-repo/docs/ops/FAIR_CLOCK.md — prefer this public file + sealed Jul 15–16 summaries for current claim practice.
What speed means
The published speed metric is first-passage train time:
- Train on the sealed recipe / pin.
- Log official val_disjoint on a fixed schedule (claim-stack: every 128 steps).
- Find the first step with
val_disjoint ≤ 3.3821. - Read
train_time_msat that row → convert to hours: T = train_time_ms / 3.6×10^6. - Compare to the public denominator:
ΔT% = 100 × (T_F1_seal - T_M) / (T_F1_seal) with T_F1_seal = 3.866 h
Positive ΔT% = fewer hours to first hit the quality bar (faster).
Include compile warmup in T. Do not strip startup, do not substitute sliding/EMA, do not use “pretty curve” screenshots.
Public denominator
| Label | Workspace | Recipe class | First-cross T |
|---|---|---|---|
| F1_seal | Clock-W | Stock AdamW · raw FineWeb · zloss 0 · no prior/EMA | 3.866 h @ step 4736 |
Source: nocap-repo/results/f1_seal/SUMMARY.md
Exact ms @ cross: 13,918,667.5 → CSV hours ≈ 3.866297; publish 3.866.
Optional same-workspace check (not a replacement denom):
| Label | Workspace | T | Note |
|---|---|---|---|
| F1_Claim-W2 | Claim-W2 | 3.858 h @4768 | ~+0.2% vs F1_seal; used for seed-2029 same-workspace ratios |
What is NOT the metric
| Not the metric | Why |
|---|---|
| Terminal val alone | Quality without first-passage T is not a speed claim. |
| step_avg alone | Forensic only; published % uses first-cross T. |
| Sliding / EMA val | Not the official disjoint endpoint. |
| Proxy @800 ΔT% | Ordinal ranking only — never published cardinal %. |
| Historical notebook F1 (~229.1 min) | Torch unlogged; provisional. |
| Matched ΔT% vs joint-QKV stock | Stock never crossed on sealed pairs → N/A. |
| “Looks faster on the dashboard” | Must recompute from CSV / sealed SUMMARY. |
Pin (required for cardinal %)
| Knob | Sealed value |
|---|---|
| GPU | NVIDIA A100-SXM4-40GB exact |
| Torch | 2.13.0 (resolved 2.13.0+cu130) |
| Quality bar | val_disjoint ≤ 3.3821 |
| Clock field | train_time / train_time_ms |
Fingerprint every run (profile/workspace code, app id, full torch string, cuda, gpu_name, seed, batch×accum×seq, kit). Seed 4242 required a relaunch after an 80GB assignment; the seal is 40GB.
Under-claiming / noise floor
Measured stack drift ~6% step-time is the empirical prior for single-seed noise.
| |ΔT%| vs F1_seal | Publish rule |
|---|---|
| < ~6% (single seed) | Do not publish as win — noise / exploratory. |
| 6–10% | Sealed + multi-seed preferred; report N^* + step_avg. |
| > 10% or ≥200 steps earlier | One sealed seed OK for narrative; still note first-passage variance. |
Claim-stack sealed candidates: +8.1% / +7.7% / +9.2% on seeds 1337 / 2029 / 4242 (all ≥6%). Mean ≈ +8.3%.
Quality vs speed
| Claim type | Needs sealed fair clock? |
|---|---|
| Hit/miss val ≤ 3.3821 | Absolute quality; still log fingerprint. |
| Cardinal ΔT% vs F1 | Yes — same GPU class + torch pin; published % uses F1_seal. |
Examples:
- F2b: quality HIT; historical wall ≈ −2.9% vs notebook F1 → retracted as speed evidence (below noise floor / stack confound).
- claim_muon_wsd: quality HIT; T ≈ 3.899 h → ~0.8% slower than F1_seal → not under-baseline.
- claim_muon_par: quality MISS → speed N/A.
Explicit non-claims
- Not an official NoCap win. Local seals ≠ organizer acceptance.
- No matched candidate-vs-stock first-cross speedup. Joint stock missed quality on 1337/2029; 4242 had no stock arm.
- No % vs historical F1.
- No board % from 800-step proxies.
- Candidate recipe ≠ joint stock. Headline compares candidate to AdamW F1_seal, not to joint Muon stock.
- Do not say “we officially won” or inflate means beyond sealed rounding (mean T ≈ 3.544 h, mean ΔT% ≈ +8.3%).
Clause-by-clause hedge defense: not-claimed-explained.html.
Defensible language (copy-safe)
On three independent seeds (1337, 2029, 4242), the claim-stack candidate (Muon B32 + timing-cheap split-QKV) first-crosses val ≤ 3.3821 about 8% faster than the sealed stock AdamW F1 baseline (3.866 h), with terminal CE ≈ 3.37. Where measured, the joint-QKV stock control misses the quality bar.
Not claimed. Not an official NoCap win. Matched cand-vs-stock ΔT% = N/A. Public % vs F1_seal only.