NoCap-Test evidence · think2earn docs
Source code ↗

Canonical numbers for the public page

Evidence state: 2026-07-16 · claim-stack v1 · 3/3 candidate seals
Resolution order: sealed *_SUMMARY.md / MULTI_SEED_CLAIM.md → pulled CSVs → this file → page copy
Do not publish % vs historical F1 (~229.1 min). Do not claim matched first-cross ΔT% vs joint stock.

Paths below are under /home/q/Downloads/software_i_built/modal/ unless noted.


Public denominator — F1_seal

FieldCanonicalSource
WorkspaceClock-Wpublic code; real profile in WORKSPACE_MAP.local.md only
GPUNVIDIA A100-SXM4-40GBsame
torch2.13.0+cu130same
First-cross step4736same + nocap-repo/results/f1_seal/log.csv
val @ cross3.381538same
train_time_ms @ cross13,918,667.5same
T (published)3.866 hSUMMARY (CSV exact ÷3.6e6 = 3.866297; publish 3.866)
Terminal val @ 47683.379813same
Historical F1~229.1 min provisional onlySUMMARY — never for published %

ΔT% formula (all published %): 100 × (3.866 − T_h) / 3.866


Same-workspace check — F1_Claim-W2

FieldCanonicalSource
WorkspaceClaim-W2public code; local file F1_BCAROL_SUMMARY.md
First-cross@4768 · val 3.380977same + CSV
train_time_ms13,889,386.9same
T (published)3.858 h (3.8582 detailed)same
vs F1_seal~+0.2%same

Candidate claim-stack (timing-cheap split-QKV)

Recipe pin (all three seeds): B32×accum16 · Muon LR 0.02 · mom 0.95 · Muon WD 0.1 · Adam LR 0.0018 · Adam WD 0.1 · NS 5 · zloss 0 · frozen unigram prior · raw FineWeb · wu 476 / wd 1192 · 4768 steps · A100-SXM4-40GB · torch 2.13 · muon_split_qkv=1 · match_joint=0.
Sources: PREREG.md, MULTI_SEED_CLAIM.md.

SeedWorkspaceCrossval @ crossT (pub)ΔT% vs F1_sealTerminal CEStockCSV recompute
1337Claim-W147363.37413.553 h+8.1%3.3710final 3.3848 · no crossno local cand CSV — sealed SUMMARY only
2029Claim-W247363.3726 (3.372588)3.567 h (3.5669)+7.7%3.3695 (3.369498)final 3.3840 (3.384017) · no cross…/seed2029/cand_log.csv · ms 12,840,827.6 → 3.566897 h · ΔT% 7.74 → +7.7%
4242Claim-W147363.3733 (3.373309)3.511 h (3.5112)+9.2%3.3702 (3.370249)not run…/seed4242/cand_log.csv · ms 12,640,470.5 → 3.511242 h · ΔT% 9.18 → +9.2%

Aggregates (from published rounded T):

AggregateValue
Mean cand T≈ 3.544 h (exact mean of 3.553/3.567/3.511 = 3.543667)
Mean ΔT% vs F1_seal≈ +8.3% (mean of +8.1/+7.7/+9.2 = +8.337…)
Seed 2029 vs F1_Claim-W2+7.5%
Bar widths vs F1 (display)91.9% · 92.3% · 90.8%

All three candidates first-cross at step 4736.


Stock control (joint QKV · same stack minus split)

SeedTerminal CEFirst-crossSource
13373.3848NONESEED1337_SUMMARY.md
20293.3840NONESEED2029_SUMMARY.md + seed2029/stock_log.csv
4242not runSEED4242_SUMMARY.md

Matched candidate-vs-stock first-cross ΔT% = N/A.


Defensible one-liner

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 our sealed stock AdamW F1 baseline (3.866 h; mean ΔT% ≈ +8.3%), with terminal CE ≈ 3.37. Where measured, the joint-QKV stock control misses the quality bar.

Attribution: Keller Jordan Muon + timing-cheap split-QKV (this campaign).
Status language: multi-seed under-baseline vs sealed F1. Under our sealed F1 baseline. Matched cand-vs-stock ΔT% = N/A. Public % vs F1_seal only.


Seed 4242 ops note (non-metric)

First spawn rejected (A100-SXM4-80GB). Sealed run is relaunched 40GB. App ap-cl3lzJm0tSNWHMdBp0LPuX · runs/claim_stack_s3_cand_seed4242_4768. Do not put secrets on the public page.