NoCap/evidence 2026-07-16 · 3/3 sealed

BottleCapAI NoCap-Test · sealed campaign report

NoCap

On three independent seeds (1337, 2029, 4242), the claim-stack candidate—Keller Jordan’s Muon with B32 plus timing-cheap split-QKV—first reaches val_disjoint ≤ 3.3821 in 3.553 h, 3.567 h, and 3.511 h. That is 8.1%, 7.7%, and 9.2% less first-passage train time than our sealed stock AdamW F1 baseline at 3.866 h (mean ≈3.544 h, mean ΔT% ≈+8.3%).

  • GPT-2 ≈124M · d12 · FineWeb
  • 1× A100-SXM4-40GB
  • torch 2.13.0+cu130
  • official disjoint val

Caveats on first screen: vs our sealed F1—not an official NoCap win. Joint-QKV stock control missed quality on seeds that ran it (1337/2029); seed 4242 had no stock arm. Matched candidate-vs-stock first-cross ΔT% is N/A. Pending organizer rules and submission.

3 / 3
seeds under F1_seal
≈8.3%
mean ΔT% vs F1
3.866 h
F1_seal T_cross
≈3.37
candidate terminal CE
Supported Candidate first-crosses ≤3.3821 on 3/3 seeds, each ≥6% under sealed F1_seal on the matched GPU/torch pin.
Not supported Matched first-cross speedup vs joint-QKV stock: stock never crossed on 1337/2029; not run on 4242.
Status Multi-seed under-baseline evidence against our seal—pending organizer rules and submission.

01 / Scoreboard

The sealed comparison

Train GPT-2 d12 on FineWeb until official disjoint validation ≤ 3.3821, faster than the sealed baseline on one GPU. “Cross” = first logged row that hits that threshold; T = train_time_ms at that row.

Public denominator: F1_seal on Clock-W. Horizontal scroll on small screens.
Run Role / workspace First passage T_cross vs F1_seal Terminal CE Status
F1_seal stock AdamW baseline denominator
Clock-W		 @4736val 3.381538 3.866 h13,918,667.5 ms reference 3.379813 sealed
candidate · s1337 Muon B32 + cheap split-QKV claim seed
Claim-W1		 @4736val 3.3741 3.553 h +8.1% 3.3710 hit
candidate · s2029 Muon B32 + cheap split-QKV claim seed
Claim-W2		 @4736val 3.3726 3.567 h +7.7% 3.3695 hit
candidate · s4242 Muon B32 + cheap split-QKV claim seed
Claim-W1 · cand-only		 @4736val 3.3733 3.511 h +9.2% 3.3702 hit

Matched stock control

Same Muon stack with joint QKV: terminal 3.3848 (s1337) and 3.3840 (s2029)—both above the bar, no first cross. Seed 4242 stock arm was not run. Matched candidate-vs-stock first-passage ΔT% is N/A.

Same-account cross-check

F1_Claim-W2 crossed @4768, val 3.380977, 3.858 h (~0.2% vs F1_seal). Seed 2029 is also 7.5% faster against that same-workspace denominator. Seed 4242’s first spawn hit an A100-SXM4-80GB guard reject; the seal is the relaunched 40GB run.

02 / How we measure

First passage, not a screenshot

Quality and speed are separate bars. Sliding/EMA curves are diagnostics only. Primary endpoint: first train_time_ms where val_disjoint ≤ 3.3821.

01 / Clock

Training time

Not calendar wall. Harness pauses the timer for validation; compile warmup stays inside the measured path.

02 / Metric

Disjoint val

Official disjoint set defines the cross. Sliding and EMA never define T.

03 / Pin

Exact hardware

A100-SXM4-40GB · torch 2.13.0+cu130 · fingerprint seed, batch, compile, software identity.

04 / Noise

~6% floor

Single-seed noise prior ≈6%. Band 6–10% wants replication; this candidate is 3/3.

Denominator rule

Published percentages use sealed F1_seal (3.866 h) on the matched GPU/torch pin. Historical F1 (~229.1 min) is provisional only—not used for any published %. Same-workspace F1_Claim-W2 is reported separately where available.

How evidence was produced
  • Kill before seal. Weak levers die at proxy cost with pre-registered stops.
  • Bank components. Systems B32 path and timing-cheap split-QKV cleared promote criteria before the full stack.
  • Seal candidate + joint control on the same pin (where both arms were scheduled).
  • Keep “no” on the page. Failures stay in the ledger.
  • Fire-and-forget ops. Long seals use detached spawn(); volumes do not share across accounts.

03 / Timeline

Mostly failed ideas, then a stack

Causal order: early waves → failed finals → fair-clock reform → failed claims → component bank → three seals.

FoundationLearn / park

Early waves found pieces—and bad extrapolations

Wave 1 on Colab T4 exposed prior, sliding-eval, and warmdown behavior. Later waves parked naive Canon (~30% step-time tax) and banked a ~1.6% parallel-block timing gain only within that notebook-era family.

First finalsMiss / retraction

Quality recovery was not a speed result

F2 (warmdown 256) ended near 3.405, never crossed. F2b (warmdown 1024) ended near 3.377 and crossed, but the −2.9% historical-wall framing was retracted (stack drift + ~6% floor).

Fair-clockMethod reset

Seal the denominator before another %

F1_seal fixed the public baseline at 3.866 h on exact A100-SXM4-40GB / torch 2.13.0+cu130. Historical F1 became provisional; single-seed noise prior ≈6%.

Full claimsNo promote

Muon’s proxy lead did not auto-survive the horizon

claim_muon_par ended ~3.386, no cross. claim_muon_wsd: quality ~3.376, but T=3.899 h ≈0.8% slower than F1_seal. Board-v3 screens: no promotion.

Component bankTwo pieces

Systems path + small quality edge

B32×accum16 ~6% faster core step vs B16 (same 524,288 tok/update). Timing-cheap split-QKV ≈+0.015 CE at 800 steps; +0.0147 CE at 2400 on seed 5151. Tier A: 5/5 no-go.

Claim stack3 / 3 sealed

Composite candidate cleared the bar three times

s1337 (Claim-W1) 3.553 h · s2029 (Claim-W2) 3.567 h · s4242 (Claim-W1, cand-only) 3.511 h. Joint controls missed on 1337/2029. Claim is candidate vs sealed F1—not matched first-cross vs joint stock.

04 / Recipe

What was actually sealed

Composite recipe. Stock control holds the stack fixed and removes only the split-QKV treatment.

claim-stack v1 / candidate

seeds 1337 · 2029 · 4242

Model & data

Model
GPT-2 d12
Data
raw FineWeb
Seq len
1,024
Val
official disjoint
Val batch
16
Prior
frozen unigram

Batch & horizon

Microbatch
32
Accumulation
16
Tok / update
524,288
Updates
4,768

Optimizer

Muon LR
0.02
Momentum
0.95
Muon WD
0.1
Adam LR
0.0018
Adam WD
0.1
NS steps
5

Schedule & loss

Warmup
476
Warmdown
1,192
Z-loss
0
EMA
off
Val every
128
Save every
256

Systems

Precision
bf16
Compile
on
Fast path
single process
Block
parallel

Environment

GPU
A100-SXM4-40GB
Torch pin
2.13.0
Resolved
2.13.0+cu130
GPU count
1

05 / Negative results

The bad news stays visible

Removing these would make the story cleaner and the evidence weaker.

Failed runs, killed components, retractions, claim boundaries. Swipe on small screens.
Item Verdict What the evidence says
F2 · short warmdown quality miss Warmdown 256 ended near 3.405; never reached 3.3821.
F2b · long warmdown speed retracted Final CE ~3.377 is a quality hit. −2.9% vs historical F1 is not speed evidence.
claim_muon_par quality miss Ended near 3.386; never crossed.
claim_muon_wsd speed miss Quality ~3.376; first passage 3.899 h ≈0.8% slower than F1_seal.
FineWeb curation screens no promote 800-update screens: no arm promoted. Sealed claim uses raw FineWeb.
Adam-group LR 0.003 no-go 800-step lead did not persist; 2400-step gates failed.
Dual-polar split-QKV time tax ~+0.015 CE but timing ratio ≈1.026 failed the ceiling.
Tier A hyperparameters 5 / 5 no-go Mom 0.90/0.98, Adam WD 0.05/0.2, warmdown 160: no promotion.
Joint-QKV stock control 2 / 2 no cross Terminals 3.3848 / 3.3840 on seeds with a stock arm. Seed 4242: not run.
Candidate vs stock ΔT% N/A No matched first-cross % — stock never crossed where measured.
Split-QKV alone not sufficient Component ≈+0.015 CE ≪ ≈0.06 CE equal-speed pure-quality bar.
Unsealed account mixing forbidden Proxies may rank methods; they do not mint published cardinal % vs an unsealed clock.

06 / Workspace map

Roles, not interchangeable clocks

Workspaces isolate volumes. Codes below are provenance labels—not a leaderboard and not public account names.

Clock

Sealed stock denominator. Cardinal reporting starts from a fingerprinted seal.

Claim pairs

Full-horizon candidate seals; matched stock where scheduled; same-workspace F1 on Claim-W2.

Research

Proxies and no-gos. Ordinal until promoted into a sealed claim.

Anonymized workspace provenance. Swipe on small screens.
Workspace code Campaign role How evidence is used
Clock-W Clock: F1_seal Public stock denominator
Claim-W1 Claim seeds 1337 + 4242 First + third candidate seals (4242 cand-only)
Claim-W2 Claim seed 2029 + F1 re-seal Second candidate seal; same-workspace check
Research-W1 Research board / Wave B Ordinal ranking only
Research-W2 Horizon screens + claim_muon_wsd Screens + full speed miss
Research-W3 / Research-W4 Cheap-split bank + Tier A Component repeat; 2400-step persistence (Research-W4)
Research-W5 Early mechanism repeats 800-step split + Adam-LR evidence
Research-W6 Incomplete 2400 attempt Archived; not a seal
Archive-W1 / Archive-W2 / Archive-W3 Historical scale / F1 / F2b History and diagnostics only

07 / Evidence paths

Public links, then local seals

Clickable URLs below are verified public. Local seal paths are labels for reviewers with the private campaign bundle—not web downloads. No credentials on this page.

Headline seals (local)

  • nocap-repo/results/claim_stack_v1/MULTI_SEED_CLAIM.md
  • …/SEED1337_SUMMARY.md
  • …/SEED2029_SUMMARY.md
  • …/SEED4242_SUMMARY.md
  • …/F1_Claim-W2 summary (local seal note)
  • nocap-repo/results/f1_seal/SUMMARY.md

Protocol & history (local)

  • nocap-repo/docs/ops/FAIR_CLOCK.md
  • …/claim_stack_v1/PREREG.md
  • …/systems_v1/SUMMARY.md
  • …/theory_split_qkv_cheap_v1/SUMMARY.md
  • …/claim_muon_par/SUMMARY.md
  • results/claim_muon_wsd/log.csv

Reviewer contract

Full replication needs private trainer/data. Log audit does not. Same endpoint either way—do not substitute a prettier curve.

  1. Confirm exact A100-SXM4-40GB and torch 2.13.0+cu130.
  2. Raw FineWeb + official disjoint val + frozen recipe above.
  3. Run the seed through 4,768 updates; keep compile warmup in T.
  4. First CSV row with val_disjoint ≤ 3.3821 → read train_time_ms.
  5. Compare % to F1_seal; also same-workspace F1_Claim-W2 where available.
  6. Report terminal CE, step, environment, and all no-cross controls.

08 / Documentation

Methods appendix (markdown)

Companion docs for strangers and agents. On a local checkout of this staging folder they open as files; on the live site they are available only after a Pages deploy that includes these .md files (production currently ships index.html only). Workspace codes only — no Modal profile names.

Architecture appendix (optional)

The public story is results-first. Architecture mattered as a delta: timing-cheap split-QKV sits on a Muon hybrid optimizer, B32×accum16 systems path, frozen prior, WSD schedule (wu476/wd1192), bf16+compile, and parallel block. Dual-polar split was killed on timing tax (~1.026). Joint QKV on the same stack is the control that quality-missed. This page is the public evidence report; private trainer/kit remain offline.