Challenge terms — official rules vs this campaign

Audience: strangers auditing whether our numbers match the BottleCapAI benchmark.
Official source: BottleCapAI/NoCap-Test (master README).
Our measurement: FAIR_CLOCK.html · GLOSSARY.html

What BottleCapAI asks for

Open benchmark (Modded-NanoGPT fork) on one GPU (reference: RTX 4090).

Official rule	Summary
Optional goal	val loss ≤ 3.3821 on FineWeb faster than stock baseline.
Non-4090	Benchmark baseline on your setup; report speedup vs that baseline.
Fair comparison	Same treatment for baseline and candidate (e.g. don’t raise LR only on candidate).
Documentation	`IDEA.md` with motivation, method, results. Negative results welcome.
Submission	`git bundle create <first>-<last>.bundle --all` → [email protected] (speedup % in subject; negative OK). Beat baseline → add `RESULTS.md`.
Status	First competition finished Nov 2025; self-paced submissions still welcome.

They want algorithmic ideas, not LR-only tuning or copied Modded-NanoGPT tricks. Sub-baseline runs with a clear idea still matter.

Metric: cross-entropy on FineWeb val (we log val_disjoint). README allows other sequence lengths if the full val set remains a valid probability model.

How this campaign maps those rules

Official	Our mapping
Quality ≤ 3.3821	First logged `val_disjoint ≤ 3.3821` (horizon 4768). GLOSSARY.
Speed vs baseline	First-passage train time T from `train_time_ms` at first-cross (compile warmup in). FAIR_CLOCK.
Non-4090 rule	Denom = our sealed stock AdamW on pinned stack — not the 4090 leaderboard row.
Fair comparison	One sealed denom; no matched speedup vs joint stock where stock missed quality.
IDEA.md	Campaign docs + FAILURE_LEDGER.html.
Bundle	Campaign packaging for organizers.

Public denominator

Label	Workspace	Pin	First-cross T
F1_seal	Clock-W	Stock AdamW · A100-SXM4-40GB · torch 2.13.0+cu130	3.866 h @ 4736

ΔT% = 100 × (3.866 h - T_candidate) / 3.866 h

Current claim (honest)

Candidate: claim-stack v1 (Muon B32 + timing-cheap split-QKV), seeds 1337 / 2029 / 4242.
vs F1_seal: 3/3 under denom — +8.1% / +7.7% / +9.2%; mean ≈ +8.3% (~3.544 h).
Under our sealed F1 baseline (not Nov 2025 prize framing).
Not matched candidate-vs-joint-stock speedup (stock missed; 4242 no stock arm).
Submission pending to [email protected].

Attribution for any submission: Muon (Keller Jordan); timing-cheap split-QKV from this campaign’s bank.

Non-claims

4090-table speed without re-benchmarking 4090 baseline.
Cardinal % from 800-step proxies or historical notebook F1 (~229.1 min).
“Beat joint stock” — stock missed on matched pairs.

Copy-safe wording: FAIR_CLOCK § non-claims.

Doc	Use
GLOSSARY.html	Terms (val_disjoint, F1_seal, workspaces)
disjoint-val-explained.html	Pedagogy: held-out val · bar · first-passage
FAIR_CLOCK.html	Speed metric, noise floor, pin
CAMPAIGN_OVERVIEW.html	Narrative
DOCS_INDEX.html	Full catalog