Defense of the public hedge · not an apology
Not claimed
Why each clause exists · under-baseline evidence vs our seal only
The intro hedge is a claim boundary, not hedging language for weak results. We publish sealed under-baseline evidence. We do not invent a matched-pair speedup the data cannot support.
Source copy lives on intro.html#hedge. Numbers: NUMBERS_CANON · clock: FAIR_CLOCK.
Three terms · brief
- First-cross / first-passage
- First logged step where
val_disjoint ≤ 3.3821. Speed uses wall-clock T at that row — not terminal CE alone. - ΔT%
- 100 × (T_denom − T_cand) / T_denom. Public published % uses denom F1_seal = 3.866 h. Matched % vs joint stock needs both arms to first-cross — otherwise N/A.
- Joint-QKV stock
- Matched claim-stack control: same Muon B32 recipe, fused (joint) QKV — not AdamW F1. Differs from candidate only on QKV treatment.
Clause 01
Not official win
Campaign evidence vs our denom. Not a leaderboard ruling.
Clause 02
vs F1_seal only
Public % denom = AdamW 3.866 h on Clock-W.
Clause 03
Matched ΔT% N/A
No honest pair speedup without a stock first-cross T.
Beat 01 · Not an official NoCap win — local seal ≠ organizer acceptance.
Three clauses, one hedge. Digits unchanged — see NUMBERS_CANON / paste block above.
Not an official NoCap win
The organizer has not accepted this as a win. We publish under-baseline evidence against our own sealed AdamW F1 — a fair-clock comparison strangers can recompute. That is campaign evidence, not a leaderboard ruling.
Local seals are campaign evidence against our own denom — not BottleCapAI acceptance.
Only under our sealed F1
Public ΔT% uses one denom: AdamW F1_seal on Clock-W (first-cross T = 3.866 h). Not historical notebook F1 (~229.1 min, provisional). Not “beat the public leaderboard.”
Under-baseline vs F1_seal is the sealed claim language. See FAIR_CLOCK for the clock and NUMBERS_CANON for the table.
Matched cand-vs-stock ΔT% = N/A
Predict · Stock never crossed — can we still publish a matched speedup?
No. Matched ΔT% needs both arms to first-cross. Without a stock T, the honest endpoint is N/A — not “infinite win,” not silent omission.
PREREG wanted matched first-cross % vs joint-QKV Muon stock (same seed, pin, recipe; only QKV treatment differs). Observed: stock never first-crossed on seeds 1337 / 2029; seed 4242 had no stock arm. Without a stock first-cross T, an honest matched speedup cannot be computed.
Published
vs F1_seal
Candidate first-cross T vs AdamW denom 3.866 h → mean ≈ +8.3%.
Not published
matched ΔT% N/A
Joint Muon stock never crossed where measured — no honest pair speedup.
Loading chart…
Predict · At final K, can matched ΔT% ever turn numeric?
No. Stock ends at 3.384 > bar 3.3821 — never first-crosses. Cand hits → published +7.7% vs F1_seal only. Matched pair stays N/A.
Pattern: TensorTonic precision-recall-at-k (K slider + live readouts) +
cross-entropy-loss (horizontal CE threshold).
Data: data/chart_series.json · s2029 cand vs joint stock.
Sealed digits: NUMBERS_CANON · F1_seal · QKV primer · timing-cheap split.
Public % is therefore vs F1_seal only. The joint miss is still useful: a quality ablation (control fails the bar where measured) — not a silent matched win. Ledger: FAILURE_LEDGER.
Related
- Intro · Not claimed hedge
- f1-seal-explained.html — denom pedagogy (3.866 h)
- seeds-explained.html — 3/3 seal bar
- FAIR_CLOCK — speed metric + non-claims
- NUMBERS_CANON — sealed T / ΔT% table
- FAILURE_LEDGER — joint miss / non-claims
- Kill → bank → 3/3 · hedges
- GLOSSARY · Not claimed / matched ΔT% N/A
- Report · Verification + hedges
- DOCS_INDEX