NoCap-Test evidence · think2earn process explainer · campaign jargon
Source code ↗

Learn the campaign process · not a model knob

Kill → bank → 3/3

Campaign process jargon · close dead ends, keep building blocks, then seal

This is process discipline, not an optimizer hyperparameter. Under the fair clock, the stack only advances after a kill, a bank, and a seal — in that order.

One-line definition

Kill = stop a dead end · Bank = keep a passed component · Seal = 3/3 seeds vs F1_seal

01 Kill 02 Bank 03 Seal · 3/3
Visual · three-beat spine process · not a % claim

Beat 01

Kill

Close the dead end. Stop spending compute.

Beat 02

Bank

Keep a promote-screen pass as a block.

Beat 03

Seal · 3/3

Compose the recipe. Three seeds vs F1_seal.

Beat 01 · Kill — Close the dead end. Stop spending compute.

Bank ≠ full claim. Seal numbers live on the report / NUMBERS_CANON — this page teaches the order. Use Play or Step to walk beats; walkthrough cards below sync via data-beat.

Problem 00 · order seal last

Wrong order fails

Sealing before kill/bank confounds the story: you cannot know which ingredient earned the clock. Ablations and baselines come first. Compose only after dead ends are closed and components are banked.

Predict · Can you seal a full recipe before banking B32?

You can run it — but the campaign rule is: bank components that passed promote screens, kill lines that miss quality or under-baseline speed, then seal 3/3 vs F1_seal. Order is process discipline, not a GPU flag.

Problem 01 · kill stop compute

Kill — close the dead end

A kill ends a line under the fair clock. Quality miss, quality without under-baseline speed, or a time-tax fail — all count. The point is to stop spending GPU on a closed path.

Not a claim. A kill is evidence of what failed. It does not mint ΔT%.

Ledger examples
  • claim_muon_par — final val 3.386130; never crossed → quality miss → speed N/A.
  • claim_muon_wsd — quality HIT (~3.376 terminal); T ≈ 3.899 h ≈ 0.8% slower than F1_seal (3.866 h) → quality without speed.
  • Dual-polar split-QKV — ~+0.015 CE, time ratio ≈ 1.026 → CE edge eaten by step tax.
  • Tier A (mom / Adam WD / warmdown screens) — 5/5 NO-GO; defaults stay.
Problem 02 · bank component ≠ claim

Bank — keep the building block

A bank keeps a component that passed promote screens. Systems path, timing-cheap split — real CE/time evidence at short horizon. Still not a full-horizon under-baseline recipe.

Bank ≠ claim. Through 2026-07-14: bank existed; 0 under-baseline full recipes yet.

What entered the bank
  • B32×accum16 — ~6% faster core step-time vs B16 path (tokens/update fixed).
  • Timing-cheap split-QKV — MECH-REPEAT @800; persistence seal Research-W4 seed 5151 @2400.
  • Closed NO-GOs stay closed: dual-polar, muon_lr 0.025, NS6, late EMA-as-train-CE, Tier A variants.
Problem 03 · seal · 3/3 vs F1_seal only

Seal · 3/3 — only then claim evidence

A seal freezes full-horizon numbers after the composed recipe runs. 3/3 means the candidate first-crosses on three seeds against sealed F1_seal (3.866 h on Clock-W) — not against joint-QKV stock.

Seed Candidate T vs F1_seal Joint stock
1337 3.553 h +8.1% never crossed · ΔT% N/A
2029 3.567 h +7.7% never crossed · ΔT% N/A
4242 3.511 h +9.2% stock arm not run · ΔT% N/A

Mean cand T ≈ 3.544 h · mean ΔT% ≈ +8.3%. Source: NUMBERS_CANON.

Hedges / non-claims

Not claimed. Not an official NoCap win. Why each clause.

Published ΔT% is vs sealed F1_seal only. Matched candidate-vs-stock first-cross ΔT% = N/A.

No published % vs historical notebook F1. Numbers on this page come from NUMBERS_CANON / FAILURE_LEDGER — invent nothing.

Lit vs process. This page teaches kill → bank → seal. It does not invent optimizer or systems papers. B32×accum and timing-cheap split enter as banked components; Keller Muon / Modded-NanoGPT stay attribution links. Sealed ΔT% lives on the report — not as a process % here.