Learn the multi-seed seal · not a new knob
Seeds 1337 · 2029 · 4242
One recipe · three RNG seeds · three independent first-cross clocks
A seed here is an RNG seed for a full-horizon claim-stack run — not a hyperparameter to tune, and not a third optimizer. Same candidate recipe. Three independent draws. Each draw gets its own first-cross T and ΔT% vs sealed F1_seal (3.866 h).
One-line definition
Seed = RNG draw for one independent full run · 3/3 = all three candidate seals pass vs F1_seal
NUMBERS_CANON RECIPE_CARD GLOSSARY · workspaces GLOSSARY · Seal / 3/3 Intro · Read more Docs
One recipe
Candidate stack
+ timing-cheap split-QKV
· 4768 steps · pin fixed
Three dice
Independent seeds
Three clocks
Own T · own ΔT%
Beat 01 · Recipe — One candidate stack. Pin fixed. Seeds do not retune it.
Same recipe, three RNG draws, three first-cross times. Digits from NUMBERS_CANON. Mean cand T ≈ 3.544 h · mean ΔT% ≈ +8.3%. Not an official NoCap win.
What a seed is
In this package, 1337, 2029, and 4242 are RNG seeds for independent full claim-stack runs. They reshuffle stochasticity (init / data order / noise), not the recipe pin.
The shared recipe stays fixed: B32×accum16 · Muon + timing-cheap split-QKV ·
4768 updates · A100-SXM4-40GB · torch 2.13 · first-cross at
val_disjoint ≤ 3.3821. See RECIPE_CARD.
Not a tune knob. Seeds are independence draws for the seal bar — not a search over lucky RNG.
Why multi-seed · 3/3
First-cross time is high variance. One seed is not a confidence interval. The process bar is 3/3: the composed candidate must first-cross on three seeds, each with ΔT% vs F1_seal ≥ 6% (noise-floor prior).
Predict · Is +7.7% on one seed enough to publish?
Not as a campaign seal. Single-seed |ΔT%| under the ~6% noise floor is not a win; even above it, the seal bar is 3/3 independent first-crosses vs F1_seal. Seed 2029’s +7.7% counts as one of three — not the whole claim.
Sealed outcome: all three pass (+8.1%, +7.7%, +9.2%). Mean ΔT% ≈ +8.3%. That freezes SUMMARY + CSV for claims. It does not mint an organizer win.
Seal / 3/3 sits after kill → bank. Bank ≠ claim until the three-seed seal.
Each seed’s own T / ΔT%
Do not collapse the three runs into one invented number on a chart without saying so. Report each seed’s first-cross T, then the mean. Public denominator is always sealed F1_seal = 3.866 h (Clock-W).
Matched pair · stock arm did not cross · summary-only (no public cand CSV).
Matched pair · stock arm did not cross · also +7.5% vs F1_Claim-W2 (not public denom).
Cand-only · no stock arm run · still counts toward 3/3 vs F1_seal.
| Seed | Workspace | Cross | T | ΔT% vs F1_seal | Stock arm |
|---|---|---|---|---|---|
| 1337 | Claim-W1 | 4736 | 3.553 h | +8.1% | no cross · ΔT% N/A |
| 2029 | Claim-W2 | 4736 | 3.567 h | +7.7% | no cross · ΔT% N/A |
| 4242 | Claim-W1 | 4736 | 3.511 h | +9.2% | not run · cand-only |
All three candidates first-cross at step 4736. Seed 1337 is summary-only (no public cand CSV). Seed 2029 may also report +7.5% vs F1_Claim-W2 — that does not replace public F1_seal. Source: NUMBERS_CANON.
Claim-W1 vs Claim-W2 homes
Volumes do not share across workspaces. Each seed’s artifacts live on a named claim workspace. Public pages name workspaces by code (Clock-W, Claim-W1, Claim-W2).
Claim-W1
Seeds 1337 (matched pair) and 4242 (candidate-only).
Claim-W2
Seed 2029 (matched pair) + same-class AdamW re-seal F1_Claim-W2 (3.858 h).
F1_seal itself lives on Clock-W. That is the sole public speed denominator.
4242 · candidate-only
Seed 4242 is a robustness seed: candidate arm only. No joint-QKV stock arm was run. It is not a matched pair.
- It still counts toward the 3/3 candidate-vs-F1_seal seal.
- It does not create a matched cand-vs-stock ΔT% (that endpoint is N/A for all three seeds).
- Do not write “three matched pairs.” Only 1337 and 2029 are matched pairs — and even there, stock never crossed.
Without overclaiming. 4242’s +9.2% is candidate vs F1_seal only. It is not evidence that a stock twin would have missed or crossed. Ops note (non-metric): first spawn hit A100 80GB and was rejected; the sealed run is the relaunched 40GB pin.
Hedges / non-claims
Not an official organizer win. Accuracy hedges live on not-claimed-explained.
Published ΔT% is vs sealed F1_seal only. Matched candidate-vs-stock first-cross ΔT% = N/A.
Single seed ≠ interval. Cand-only 4242 ≠ matched pair. Numbers from NUMBERS_CANON / RECIPE_CARD — invent nothing.
Related
- GLOSSARY · Claim-W1 / Claim-W2
- GLOSSARY · Seal / 3/3
- GLOSSARY · Matched pair
- kill-bank-seal-explained.html — process spine that ends in 3/3
- report.html — sealed claim surface
- Intro · Read more
- Docs index