NoCap-Test evidence · think2earn 2026-07-16 · 3/3 sealed
Source code ↗

BottleCapAI NoCap-Test · sealed campaign report

≈8.3% less first-passage train time than F1_seal (3.866 h) — 3/3 seeds

Carol Calin · [email protected]

Applied Jul 9. Sealed Jul 15–16.

On three independent seeds (1337, 2029, 4242), the claim-stack candidate — Keller Jordan Muon with B32 plus timing-cheap split-QKV — first-crosses official disjoint val ≤ 3.3821 in 3.553 h, 3.567 h, and 3.511 h.

That is +8.1%, +7.7%, and +9.2% versus our sealed stock AdamW F1 baseline at 3.866 h (F1_seal @4736, val 3.381538). Mean ΔT% ≈ +8.3%. Terminal candidate CE ≈ 3.37.

Where measured, the joint-QKV stock control never crosses (1337, 2029). Seed 4242 stock was not run.

Pin: 1× A100-SXM4-40GB · torch 2.13.0+cu130. Workspaces: Clock-W, Claim-W1, Claim-W2.

Under our sealed F1 baseline. Public % vs F1_seal only. Matched cand-vs-stock ΔT% = N/A. Why each clause.

The metric

First-passage train time. Quality gate and speed clock are separate. The official quality gate is val_disjoint ≤ 3.3821. The speed clock is measured train_time_ms (validation paused; compile warmup inside the path).

T = train_time_ms at the first logged row where val_disjoint ≤ 3.3821, converted to hours.

T_h = train_time_ms ÷ 3,600,000
ΔT% = 100 × (3.866 − T_h) / 3.866

Denominator: sealed F1_seal on Clock-W — first-cross @4736, val 3.381538, T = 3.866 h. Historical F1 (~229.1 min) is provisional only and is never used for published %.

Why ΔT% vs F1_seal. First-passage = gate then clock: first row val_disjoint ≤ 3.3821, then train_time_ms → T_h. Public ΔT% always divides by sealed F1_seal (3.866 h, Clock-W) — not joint-QKV stock (matched first-cross ΔT% = N/A) and not historical notebook F1. Verify-recompute repeats that seal denominator: ΔT% = 100 × (3.866 − T_h) / 3.866. Campaign GPU tiers and the Muon primer are not claim speed.

Challenge rule (NoCap-Test): reach val ≤ 3.3821 on 1 GPU. If not on a 4090, seal your own baseline and compare to that. This page compares to our seal.

What the numbers say

Three candidate seals. One sealed baseline. Mean first-cross ≈ 3.544 h. Mean ΔT% ≈ +8.3%.

Run Workspace First-cross T vs F1_seal Terminal CE
F1_sealstock AdamW · denominator Clock-W @4736val 3.381538 3.866 h ref 3.379813
candidate · s1337Muon B32 + split-QKV · public CSV Claim-W1 @4736val 3.3741 3.553 h +8.1% 3.3710
candidate · s2029Muon B32 + split-QKV Claim-W2 @4736val 3.3726 3.567 h +7.7% 3.3695
candidate · s4242Muon B32 + split-QKV · cand-only Claim-W1 @4736val 3.3733 3.511 h +9.2% 3.3702

Loading first-passage chart from data/chart_series.json…

1 · Gate

Horizontal line: val ≤ 3.3821. Quality first. (= recompute step 1)

2 · Clock

Vertical F1_seal: sealed baseline 3.866 h. (= ΔT% denominator)

3 · Cross

Magenta markers: first-cross before baseline → ΔT% vs F1_seal (not vs stock).

Verify skim: gate → clock → cross mirrors #recompute (1 first-cross ≤ 3.3821 · 2 T_h = ms÷3.6e6 · 3 ΔT% vs 3.866). Muon is part of the recipe. Published ΔT% is for the full claim-stack, not Muon alone. Campaign GPU-tier totals are not the claim clock.

Loading…
Figure 1 — first-passage. Disjoint validation loss (y, lower better) vs training-time hours (x). Dashed line: target 3.3821. Dotted vertical: F1_seal first-cross at 3.866 h. Magenta marks candidates. Seed 1337 from public CSV (s1337_cand.csv). Seed 2029 joint-QKV stock never crosses. Published ΔT% is vs F1_seal only. Matched cand-vs-stock ΔT% = N/A.

Verification status

Claimed

Candidate first-crosses ≤3.3821 on 3/3 seeds. Each seed is at least 6% under F1_seal; mean ΔT% is approximately +8.3%. Recompute from public CSVs in data/.

Not claimed

Under our sealed F1 baseline. Public % vs F1_seal only. Matched cand-vs-stock ΔT% = N/A. Stock never crossed where run; 4242 stock was not run.

Chart steps (gate → clock → cross) match the formula below. Full report verify: report.html#verification.

Negative results

Item Verdict Evidence
Joint-QKV stock · s1337 no cross Terminal 3.3848. Did not first-cross val_disjoint ≤ 3.3821.
Joint-QKV stock · s2029 no cross Terminal 3.3840. Min val 3.384017.
Joint-QKV stock · s4242 not run Cand-only seal.
Candidate vs stock ΔT% N/A No matched first-cross %.
claim_muon_par quality miss Ended ~3.386. Never crossed.
claim_muon_wsd speed miss Quality ~3.376; T=3.899 h ≈0.8% slower than F1_seal.
F2 short warmdown quality miss Ended ~3.405. Never crossed.
Dual-polar split-QKV time tax ~+0.015 CE but timing ratio ≈1.026 failed ceiling.

Full ledger: FAILURE_LEDGER.html.

Recipe

Claim-stack v1 candidate. Seeds 1337, 2029, 4242. Stock control = same stack with joint QKV (muon_split_qkv=0).

Model / data
GPT-2 d12 · raw FineWeb · seq 1024 · official disjoint val
Batch
B32 × accum16 · 524,288 tok/update · 4768 steps
Muon
LR 0.02 · mom 0.95 · WD 0.1 · NS 5
Adam (embed/lm_head)
LR 0.0018 · WD 0.1
Schedule
warmup 476 · warmdown 1192 · zloss 0 · frozen unigram prior
Candidate delta
muon_split_qkv=1 · match_joint=0
Environment
A100-SXM4-40GB · torch 2.13.0+cu130 · 1 GPU · bf16 · compile on
Attribution
Keller Jordan Muon + campaign timing-cheap split-QKV

Pin card: RECIPE_CARD.html. Canon: NUMBERS_CANON.html. Muon design diagrams: muon-explained.html.

How to recompute

Verify in five steps. Gate → clock → cross on public CSVs; then stock and separate-clock hedges. Same order as Figure 1.

1 · Gate · first-cross = first CSV row where val_disjoint ≤ 3.3821
2 · Clock · T_h = train_time_ms ÷ 3.6e6
3 · Cross · ΔT% = 100 × (3.866 − T_h) / 3.866
4 · Stock · no cross → matched ΔT% = N/A
5 · Separate · see report #compute (72.03 / 110.18 / 147.91) · Muon primer

Worked example — s2029_cand.csv: step 4736, train_time_ms = 12,840,827.6 → T_h = 3.5669 → ΔT% = +7.7%.