NoCap-Test evidence · think2earn docs
Source code ↗

NoCap dev-page fact-check checklist (3/3)

Checklist for every scientific numeric claim in cf-pages-nocap-agent-dev/index.html (2026-07-17 jury-ready pass). Repeated values share one row. CSS/layout dimensions are out of scope.

Canonical table: NUMBERS_CANON.md (same folder).
Base path: /home/q/Downloads/software_i_built/modal/

Recompute date: 2026-07-17. CSV first-cross = first row with val_disjoint ≤ 3.3821; T_h = train_time_ms / 3.6e6; ΔT% = 100 × (3.866 − T_h) / 3.866.


Jury-pass status (2026-07-17 QA)

GateStatus
OverallPASS — numbers match NUMBERS_CANON.md; no public Modal-profile leaks; claim language clean
Number diff index.html ↔ canonPASS — F1_seal 3.866 h; seeds 3.553 / 3.567 / 3.511; ΔT% +8.1 / +7.7 / +9.2; mean ≈+8.3% / ≈3.544 h; stock no-cross / N/A
CSV recompute (public data/*)PASS — F1 / s2029 / s4242 / F1_Claim-W2 / s2029 stock no-cross
Modal profile leak scanPASS on deployable tree — profiles only in WORKSPACE_MAP.local.md (must-NOT-deploy); mailto [email protected] OK
Byline / Verify / chart / terms / reviewersPASS — Carol Calin + mailto; #verify + chart_series.json/CSV fetch; challenge terms; #next reviewers; CHALLENGE_TERMS.md linked
Forbidden claimsPASS — no matched stock ΔT% speed claim; Public % vs F1_seal only

Deploy exclude: WORKSPACE_MAP.local.md (must-NOT-deploy), EMAIL_SNIPPET.md (operator-only; prefer exclude).


Jury-ready additions (2026-07-17)

Claim on pageVerified?Source
Public data/f1_seal.csv matches F1_seal first-crossdata/f1_seal.csv ms 13,918,667.5 → 3.866297 h; same as nocap-repo/results/f1_seal/log.csv pull
Public data/s2029_cand.csvms 12,840,827.6 → 3.566897 h · +7.7%
Public data/s2029_stock.csv — no crossterminal val 3.384017; no row ≤ 3.3821
Public data/s4242_cand.csvms 12,640,470.5 → 3.511242 h · +9.2%
Public data/f1_claim_w2.csvms 13,889,386.9 → 3.858163 h (~0.2% vs F1_seal)
Seed 1337 labeled summary-only (no public CSV)SEED1337_SUMMARY.md; no data/s1337_*.csv
First-passage chart sourced from same data/*.csvVerify section fetches same-origin CSVs; optional data/chart_series.json is derived from those files
Author byline Carol Calin + mailto contact on pagejury-ready hero; no Modal profile names
Challenge terms: val ≤ 3.3821 ↔ first-passage train_time_msNoCap-Test README + FAIR_CLOCK + measure section
Challenge terms: non-4090 → seal baseline on your GPU ↔ F1_seal A100-SXM4-40GBNoCap-Test README + f1_seal SUMMARY
Challenge terms: 1 GPU ↔ attested pinhero pins + scoreboard
Challenge terms: fair comparison ↔ same pin + documented recipeRECIPE_CARD / PREREG + recipe section
Reviewer section: applied Jul 9; seals Jul 15–16reviewer block + caveats
Compressed workspace table — codes only, honest isolation framingworkspace section; real profiles only in WORKSPACE_MAP.local.md

Evidence state & challenge

Claim on pageVerified?Source
Evidence state 2026-07-16 · 3/3 sealednocap-repo/results/claim_stack_v1/MULTI_SEED_CLAIM.md
GPT-2 ≈124M · d12 · FineWebchallenge framing / production + README
Quality bar val_disjoint ≤ 3.3821f1_seal/SUMMARY.md; FAIR_CLOCK; all claim-stack seals
One GPUPREREG / challenge

Headline / hero

Claim on pageVerified?Source / recompute
Seeds 1337, 2029, 4242MULTI_SEED_CLAIM + three SEED*_SUMMARY
T = 3.553 / 3.567 / 3.511 hsealed summaries; CSV for 2029 & 4242 (below)
ΔT% = +8.1% / +7.7% / +9.2%MULTI_SEED; recompute from rounded T vs 3.866
F1_seal 3.866 hf1_seal/SUMMARY.md; CSV ms 13,918,667.5
Mean T ≈3.544 hmean(3.553,3.567,3.511)=3.543667
Mean ΔT% ≈+8.3%mean(+8.1,+7.7,+9.2)=+8.337…
Terminal CE ≈3.37terminals 3.3710 / 3.3695 / 3.3702
3 / 3 seeds under F1_sealMULTI_SEED
Each ≥6% under F1all three ≥7.7%
Public % vs F1_seal only; matched cand-vs-stock ΔT% = N/Aclaim language
Matched cand-vs-stock ΔT% N/Astock no-cross / not run
GPU A100-SXM4-40GB · torch 2.13.0+cu130f1_seal SUMMARY

Scoreboard rows

F1_seal

ClaimSource
@4736 · val 3.381538SUMMARY + nocap-repo/results/f1_seal/log.csv
3.866 h · 13,918,667.5 msSUMMARY (CSV ÷3.6e6 = 3.866297; publish 3.866)
Terminal 3.379813SUMMARY + CSV

Candidate s1337

ClaimSource
@4736 · val 3.3741 · T 3.553 h · +8.1% · term 3.3710SEED1337_SUMMARY.md; MULTI_SEED
Workspace Claim-W1same; real profile in WORKSPACE_MAP.local.md
Local cand CSV⚠️ absentrely on sealed SUMMARY (see ACCURACY_NOTES)

Candidate s2029

ClaimSource / CSV
@4736 · val 3.3726 · T 3.567 h · +7.7% · term 3.3695SUMMARY; CSV first-cross val 3.372588, ms 12,840,827.63.566897 h, ΔT% 7.74+7.7%; term 3.369498
Workspace Claim-W2SUMMARY; real profile in WORKSPACE_MAP.local.md
CSV pathnocap-repo/results/claim_stack_v1/seed2029/cand_log.csv

Candidate s4242

ClaimSource / CSV
@4736 · val 3.3733 · T 3.511 h · +9.2% · term 3.3702SUMMARY; CSV val 3.373309, ms 12,640,470.53.511242 h, ΔT% 9.18+9.2%; term 3.370249
cand-only / no stock armSEED4242_SUMMARY.md
CSV pathnocap-repo/results/claim_stack_v1/seed4242/cand_log.csv

Bar widths (display-only)

WidthRatio
91.9%3.553/3.866
92.3%3.567/3.866
90.8%3.511/3.866

Stock control & F1_Claim-W2 notes

Claim on pageSource / CSV
Stock s1337 terminal 3.3848 · no crossSEED1337_SUMMARY
Stock s2029 terminal 3.3840 · no crossSUMMARY; CSV terminal 3.384017, no row ≤3.3821
Seed 4242 stock not runSEED4242_SUMMARY / MULTI_SEED
Failure ledger “2 / 2 no cross” (paired stocks)correct — not a stale “2/2 seeds sealed” claim
F1_Claim-W2 @4768 · val 3.380977 · 3.858 hlocal F1_BCAROL_SUMMARY; CSV ms 13,889,386.9 → 3.858163
~0.2% vs F1_sealSUMMARY
Seed 2029 7.5% vs F1_Claim-W2SUMMARY; recompute ≈7.55%
4242 first spawn A100-80GB reject; seal is 40GBSEED4242_SUMMARY

Fair-clock / methodology

ClaimSource
Primary = first-passage train_time_ms at val≤3.3821FAIR_CLOCK; PREREG
Sliding/EMA do not define TFAIR_CLOCK
~6% single-seed noise floor; 6–10% wants multi-seedFAIR_CLOCK; f1_seal SUMMARY
Historical F1 ~229.1 min provisional onlyf1_seal SUMMARY
Published % use F1_seal 3.866 onlyhard rules + MULTI_SEED

Timeline / failure ledger (campaign history)

ClaimSource
Canon ~30% step-time tax parkedLESSONS / README
Parallel ~1.6% (notebook-era family)LESSONS / FAIR_CLOCK
F2 WD256 ended ~3.405 · no crossevidence_ledger / finals
F2b WD1024 ended ~3.377 · crossed; −2.9% historical speed retractedf2b PULL_SUMMARY + FAIR_CLOCK
claim_muon_par ~3.386 · no crossclaim_muon_par/SUMMARY
claim_muon_wsd ~3.376 · T 3.899 h0.8% slower than F1claim_muon_wsd log / board_v3
B32×accum16 ~6% faster core step vs B16; 524,288 tok/updatesystems_v1 / STACK
Cheap split-QKV ≈+0.015 CE @800; +0.0147 @2400 seed 5151STACK / p2400 SUMMARY
Tier A 5/5 no-go (mom 0.90/0.98, Adam WD 0.05/0.2, WD 160)TIER_A_SEAL
Curation 800 screens · no promote · claim uses raw FineWebSCREEN800_SUMMARY
Adam LR 0.003 · 800 lead / 2400 gates failedadam_lr persistence SUMMARY
Dual-polar ~+0.015 CE · timing ≈1.026 failedtheory_split_qkv SUMMARY
Split alone ≪ ~0.06 CE equal-speed quality barSTACK / ledger

Recipe card

ClaimSource
GPT-2 d12 · raw FineWeb · seq 1024 · val batch 16 · frozen unigramPREREG / kit
B32 × accum16 · 524,288 tok · 4,768 updatesPREREG
Muon 0.02 · mom 0.95 · Muon WD 0.1 · Adam 0.0018 · Adam WD 0.1 · NS 5PREREG
wu 476 · wd 1,192 · zloss 0 · EMA off · val every 128 · save every 256PREREG / MULTI_SEED
bf16 · compile on · single process · parallel blockPREREG / kit
Candidate: muon_split_qkv=1, match_joint=0PREREG
Stock = same stack, joint QKVMULTI_SEED

Workspaces / paths / next

ClaimNotes
Clock-W = F1_seal clockreal profile in WORKSPACE_MAP.local.md only
Claim-W1 = seeds 1337 + 4242
Claim-W2 = seed 2029 + F1_Claim-W2 re-sealpublic label was F1_Claim-W2
Evidence paths list sealed SUMMARY filesmatch NUMBERS_CANON
Optional stock arm on 4242 only if demandednot “optional 3rd candidate seed” — 3rd cand already sealed
Public % vs F1_seal only; matched cand-vs-stock ΔT% = N/Afooter + hedges
Public HTML has no Modal profile real namescodes only; see WORKSPACE_MAP.local.md

Stale-language scan (this index.html)

PatternFound?
“2/2” as sealed candidate countNo (only “2 / 2 no cross” for paired stock arms — correct)
“optional 3rd seed” for the claimNo (optional stock arm only)
mean 3.560 h / 7.9% (old 2-seed mean)No
Published % vs historical F1 / 229.1No (229.1 marked provisional only)
Matched stock first-cross ΔT% as a speed claimNo (explicitly N/A)

Numeric mismatches requiring frontend fix: none found against sealed 3/3 sources and CSV recomputes for 2029/4242/F1.

2026-07-17 pre-deploy QA: JURY PASS (see Jury-pass status above). No deploy performed in that QA pass.


ROUND_20260717_pilot (L6 merge · 2026-07-17)

Copy/structure merges into report.html / muon-explained.html only (index.html untouched). Deploy=no. Codes only.

ChangeVerified?Source / note
#compute triple 72.03 / 110.18 / 147.91 h + 128.28 NoCap-namedGPU_HOURS_LOCAL.md · GPU_HOURS_VOLUMES.local.md · GPU_HOURS_BILLED.local.md
Ratios 1.34× / 1.16× / 2.05×; notebook ~19.63 hsame billed note
First paragraph states train_time_ms ≠ billedL3 paste
#insight-kill-bank-seal: F1_seal 3.866 h; mean ΔT% ≈ +8.3%; wsd ~0.8% slower; matched ΔT% N/ANUMBERS_CANON.md · ledger / claim_muon_wsd
#numbers lead: ≥6% under F1; published % vs F1_seal only; matched ΔT% N/Acanon hedges
#failures hard non-claim (NL)same non-claim; stock no-cross / 4242 not run
#verification bridge to #compute + Muon primerinventory clocks ≠ claim ΔT%
muon-explained.html claim-boundary: NS×5 = Keller Ortho; sealed speed = full stack on report — not Muon-alone %L5; no sealed ΔT% moved onto Muon page
L1 anim-review-fp-pilot.html✅ review-onlyAccept demo; port to preview/report deferred
L2 shell 720 + sticky docnavCSS/chrome only; no claim numbers
Modal profile names on public HTML✅ nonecodes only; mailto author OK

Numeric mismatches from this round: none vs canon / GPU_HOURS.


ROUND_20260717_r2 (L6 merge · 2026-07-17)

Copy/structure into preview-schema.html / report.html / muon-explained.html (+ CSS skim). index.html untouched. Deploy=no. Codes only. English public copy only.

ChangeVerified?Source / note
Schema Figure 1 teach-beat gate → clock → crossTHRESH/BASE/S1337 from JSON/canon; legend ΔT% +7.7 / +9.2 / +8.1 unchanged
#insight-clock-denom: F1_seal 3.866 h; matched ΔT% N/A; no historical-F1 %NUMBERS_CANON.md
Verify-in-5 / #recompute: 3.3821 · 3.866 · s2029 → +7.7%same worked example; stock N/A; compute tiers separate
Report #compute tiers 72.03 / 110.18 / 147.91✅ untouchedL3/L6: no rewrite this round
Muon #split-vs-ortho: split ≠ Ortho/NS; sealed % = full stack on reportL5; no Muon-alone sealed %
Modal profile names on public HTML✅ nonecodes only; mailto author OK
Nederlands on merged public surfaces✅ noneEN voice only

Numeric mismatches from this round: none vs canon.


Public kit + TRAINER_DELTA (2026-07-17)

ChangeVerified?Source / note
kit/train_gpt2_split_qkv_cheap_v1.py SHA becfce38…ac9cbcPREREG + Claim-W1/W2 volume kit_override pull
kit/muon_opt_split_qkv_cheap_v1.py SHA 9086531f…6b1b20same
TRAINER_DELTA.md / .htmlflags muon_split_qkv=1 · match_joint=0; no new sealed %
Intro / report / RECIPE / DOCS_INDEX link kitpublic paths only

Muon-class jargon explainers (2026-07-17)

ChangeVerified?Source / note
New qkv-explained.html (Query/Key/Value primer)no sealed %; points to split / F1_seal / report
Upgraded f1-seal-explained.html worked ΔT% tablecanon 3.866 / +8.1/+7.7/+9.2 / ≈+8.3%
Upgraded split-qkv-explained.html + QKV pointerdual-polar 1.026 · flags unchanged
Upgraded not-claimed-explained.html side-by-side + figmatched ΔT% N/A hedge only
Wired intro / report / GLOSSARY / DOCS_INDEXcodes only; no new sealed numbers

Prior checklist staleness

The previous FACT_CHECK_CHECKLIST.md in this folder still described 2/2 seals and mean ≈3.560 h / ≈7.9%. That is obsolete; this file supersedes it.