NoCap-Test evidence · think2earn fair-clock explainer · denom only
Source code ↗

Learn the public denominator · fair clock

F1_seal

Sealed stock AdamW · Clock-W · first-passage T = 3.866 h · sole public ΔT% denom

F1_seal is the sealed stock AdamW baseline on workspace Clock-W and first-crosses val_disjoint ≤ 3.3821 at step 4736 with published T = 3.866 h. Every published candidate ΔT% uses this number as the denominator. It is neither the Muon claim candidate nor historical notebook F1.

AdamW stock Clock-W 3.866 h @4736 public denom

FAIR_CLOCK NUMBERS_CANON GLOSSARY · F1_seal data/f1_seal.csv intro

Definition

Speed on this site means first-passage train time: the train_time_ms value at the first logged row where val_disjoint ≤ 3.3821. F1_seal fixes that denominator for stock AdamW.

Public denominator F1_seal

Recipe: stock AdamW · raw FineWeb · zloss 0 · no prior/EMA
Workspace: Clock-W · GPU: A100-SXM4-40GB · torch 2.13.0+cu130
First-cross @ step 4736 · val 3.381538
train_time_ms = 13,918,667.5 → T = 3.866 h (publish)

Exact hours from CSV: 13,918,667.5 ÷ 3.6×106 = 3.866297; published rounding is 3.866. Source: results/f1_seal/SUMMARY.md / data/f1_seal.csv.

Problem 01 · first-passage first logged cross

What first-passage is

Train on a fixed pin. Log val_disjoint on schedule (claim-stack: every 128 steps). Find the first row with val_disjoint ≤ 3.3821. Read train_time_ms there → hours. That is T. Include compile warmup.

Predict · Is F1_seal the Muon claim candidate?

No. F1_seal is sealed stock AdamW on Clock-W. Candidates live on Claim-W* and publish ΔT% against this denom.

Visual · First logged quality cross Our commentary

Loading F1_seal curve from data/chart_series.json…

1 · Gate

Horizontal bar: first row with val_disjoint ≤ 3.3821.

2 · Clock

Read train_time_ms there → T_h. Sealed denom = 3.866 h.

3 · Cross

F1_seal first-cross @4736 fixes the public ΔT% denominator (not terminal val alone).

Loading…

Figure · Schematic. Sealed F1_seal cross is @4736; candidates on claim-stack also cross @4736 where sealed.

Problem 02 · seal freeze what is locked

What F1_seal freezes

A seal writes SUMMARY + CSV so the denominator cannot drift under later stories. F1_seal locks stock AdamW on Clock-W — not Muon, not split-QKV, not a candidate arm.

T published

3.866 h

Cross step

4736

Workspace

Clock-W

Optimizer

AdamW stock

Visual · Seal box on Clock-W Our commentary
CLOCK-W · SEALED F1_seal AdamW · 3.866 h raw FineWeb · zloss 0 no prior / no EMA denom CLAIM-W1 / W2 Muon candidates B32 + timing-cheap split publish ΔT% vs F1_seal ≠ F1_seal itself

Optional same-workspace AdamW check: F1_Claim-W2 = 3.858 h (~+0.2% vs F1_seal). Useful for seed-2029 ratios — not a replacement public denom.

Problem 03 · ΔT% candidate vs denom

How candidate % uses it

For a candidate first-cross time TM, published percent is always against F1_seal:

Formula Public ΔT%

ΔT% = 100 × (3.866 − TM) / 3.866
Positive = less first-passage train time than F1_seal

Visual · Denominator bar vs candidate Our commentary
FIRST-PASSAGE T (hours) F1_seal 3.866 h · denom cand TM shorter → positive ΔT% Sealed claim-stack means ≈ +8.3% vs this denom. See report / NUMBERS_CANON · digits unchanged.

Noise floor ~6% single-seed: |ΔT%| below that is not published as a sealed under-baseline claim. Full sealed table: NUMBERS_CANON.

Arm T (h) ΔT% vs F1_seal Note
F1_seal (denom) 3.866 — AdamW · Clock-W · @4736
Cand seed 1337 3.553 +8.1% Muon B32 + timing-cheap split
Cand seed 2029 3.567 +7.7% same recipe
Cand seed 4242 3.511 +9.2% same recipe · no stock arm
Mean (3 seeds) ≈3.544 ≈+8.3% public headline aggregate

Worked digits from NUMBERS_CANON. Example: seed 1337 → 100 × (3.866 − 3.553) / 3.866 ≈ +8.1%. Multi-seed seal: seeds-explained.html.

Claim boundary: Headline compares candidate (Muon B32 + timing-cheap split-QKV) to AdamW F1_seal — not to joint-QKV Muon stock. Stock never crossed where measured. Under our sealed F1 baseline. Public % vs F1_seal only. Matched cand-vs-stock ΔT% = N/A. Fair-clock rules: FAIR_CLOCK (no separate fair-clock-explained.html — this page + FAIR_CLOCK cover the denom).
Problem 04 · non-claims do not redefine

Hard non-claims

F1_seal is the sealed public denominator. Treat these as hard stops, not soft hedges.

Not the Muon candidate

F1_seal is stock AdamW. The claim arm is Muon + B32 + timing-cheap split on Claim-W*.

Not historical notebook F1

~229.1 min provisional; torch unlogged. Never use for published %. Only sealed 3.866 h.

Not matched stock ΔT%

Joint-QKV Muon stock did not first-cross on 1337/2029; 4242 had no stock arm. Matched cand-vs-stock ΔT% = N/A.

Not proxy / EMA / terminal-only

800-step proxies rank only. Sliding/EMA and terminal CE alone do not define published speed.

Full checklist: not-claimed-explained.html · FAIR_CLOCK · explicit non-claims · FAILURE_LEDGER.