Campaign timeline — causal order

Evidence state: 2026-07-16. Prefer sealed Jul 15–16 claim-stack summaries over stale README / early FAIR_CLOCK wording.

Short sentences. Forensic, not promotional.

1. Early waves → failed finals

Notebook- and archive-era work established the absolute quality bar: disjoint val ≤ 3.3821 on GPT-2 ~124M / FineWeb, one A100.

F2 (short warmdown 256, curated/prior/EMA stack) never crossed. Terminal CE ≈ 3.405. Quality miss. See nocap-repo/results/evidence_ledger_v1/SUMMARY.md (f2_record).

F2b (warmdown 1024, same family) crossed at step 4736 with terminal CE ≈ 3.377. Quality hit. Historical wall framing of about −2.9% vs notebook F1 was retracted as speed evidence (infra drift / stack confound; below the ~6% publish floor). Keep the quality hit; do not cite the old speed %. See docs/ops/FAIR_CLOCK.md and results/finals/f2b_record/PULL_SUMMARY.md.

2. Fair-clock reform → sealed F1 denominator

Stack pin: A100-SXM4-40GB · torch 2.13.0+cu130. Metric T = train_time at first logged disjoint val ≤ 3.3821.

F1_seal on Clock-W: stock AdamW, raw FineWeb, no prior/EMA. Cross @4736 · val 3.381538 · T=3.866 h. This is the public speed denominator. See results/f1_seal/SUMMARY.md.

Historical notebook F1 (~229.1 min) is provisional only (torch unlogged). Do not use it for published %.

3. Failed full claims (pre–component bank)

claim_muon_par (Clock-W): Muon + parallel + triage + prior + EMA. Full 4768. Final val 3.386130. Never crossed. Quality miss → no speed claim. Proxies @800 did not transfer. See results/claim_muon_par/SUMMARY.md.

claim_muon_wsd: quality HIT (cross @4736, terminal ≈ 3.376) but T ≈ 3.899 h — about 0.8% slower than F1_seal. Quality without speed. Logs live outside nocap-repo/ at workspace results/claim_muon_wsd/; board writeup in results/board_v3/WRITEUP.md.

Board-v3 800-step screens after that run: no promote.

4. Component bank (research accounts)

Systems path B32×accum16: ~6% faster core step-time vs B16 path (matched systems_v1). Tokens/step stay 524,288.

Timing-cheap split-QKV (block NS5, match_joint=0): multi-seed MECHANISM-REPEAT @800 · mean terminal gain ≈ +0.015 CE · time ratios ≤ ~1.01. Persistence sealed @2400 on Research-W4 seed 5151 · +0.0147 CE · no late reverse.

Many other levers closed as NO-GO (curation cleaners, Adam LR 0.003 alone, dual-polar split, muon_lr 0.025, interleave, NS6, warmdown 300, late EMA train CE, etc.). See results/evidence_ledger_v3/SUMMARY.md and results/claim_ready_stack_v1/STACK.md.

As of 2026-07-14 evening: under-baseline complete recipes still 0. Bank existed; full claim did not.

5. Tier A hparam screens — all NO-GO

Mid-horizon (800-step) screens on Research-W3 / Research-W4. Zero promote.

Defaults kept: mom 0.95, Adam WD 0.1, warmdown 200 (proxy; full claim uses wu476/wd1192). See results/research_atomic_v1/TIER_A_SEAL.md.

6. Claim-stack seals (3/3 candidates under F1)

Recipe: Muon B32 systems path + timing-cheap split-QKV candidate vs joint-QKV stock control. Full 4768. Exact A100-40 + torch 2.13. Fire-and-forget spawn() after nested .remote() cancels killed earlier work.

All three candidates first-cross at step 4736. Mean cand T ≈ 3.544 h · mean ΔT% vs F1_seal ≈ +8.3%.

Same-account AdamW re-seal F1_Claim-W2 on Claim-W2: T=3.858 h (~+0.2% vs F1_seal). Supports seed-2029 same-account ratios.

Seed 4242 ops note: first spawn failed GPU guard (Modal gave A100-SXM4-80GB). Relaunch got 40GB. Seal is the 40GB run.

Do not claim matched ΔT% vs joint stock. Stock never crossed where measured (1337, 2029); 4242 has no stock arm.

Source of truth: results/claim_stack_v1/MULTI_SEED_CLAIM.md.

Causal spine (one line)

Failed finals → fair clock + F1_seal → Muon compose misses → bank systems + cheap split → Tier A keeps defaults → claim-stack 3/3 under sealed F1 (candidate recipe, not matched stock speedup).