SKILL.md
Verify
Sample N isolated trajectories, then rank them (Kwok et al., arXiv:2607.05391). One score token is an LM judge, not this method.
Flags
| Flag | Default | Effect |
|---|---|---|
--candidates <n> |
3 |
Pool size N. Not inferred from the task |
--evals <k> |
2 |
Repeats K; swap A/B on every odd pass |
--pivots <k> |
2 |
PPT pivots when N>3; clamp to [1, N] |
--max-rounds <n> |
0 |
Extra generate or revise cycles after select |
--criteria <path> |
auto | Criteria file; else write 2–4 inline |
--track |
off | Progress only; skip generate and select |
A supplied list overrides --candidates. N is a budget, not a computed count.
1. Lock task and criteria
Require a task. Write 2–4 independent criteria plus a ground-truth note that trusts observed tool output, not narration. Write that file using the layout in REFERENCE.md.
Record: task | criteria | generate | path | n/k/pivots/rounds | current | terminal
Done when task and criteria are fixed, or BLOCKED.
2. Generate
--track: skip. If the user supplied attempts, use them unchanged.
Else dispatch --candidates workers in one parallel wave. Each brief: locked task + criteria + "one complete attempt; no siblings; no rank." Write scope is a worktree or copy using Isolation in REFERENCE.md, never a branch on this tree. Same-tree parallel writes are BLOCKED. Parent implements nothing. No spawn: N serial attempts (generate=serial).
Done when N traces are locked, or BLOCKED.
3. Verify
Run candidates when a criterion is empirical. Classify all-pass / all-fail / swing. Skip scoring on skip classes. all-fail is unwinnable here: if --max-rounds remain, consume one and return to Step 2; else ALL_FAIL.
On swing, one criterion per comparison. N≤3: every directed pair. N>3: clamp --pivots to [1, N], then run the PPT procedure in REFERENCE.md. Average over C and K. Soft win p = 1/(1+exp(-(Ra-Rb))). Accumulate wi, ci.
Scoring path (first match); pick it from REFERENCE.md:
- Logprobs at
<scoreA>/<scoreB>:R = E[φ(v)]on A–T (paper). - No logprobs, user named a logprob verifier: two-stage (paper B.6).
- This harness only: emit A–T, map, average over K and C. Not Eq. 3.1.
Still pairwise, decomposed, swapped. Never a 1–10.
--track: score the prefix against "already complete?"; no siblings.
Done when each scored candidate has wi/ci, or a skip class applies.
4. Select
--track or all-fail: no winner. all-pass: lowest-index passer. swing: argmax wi/ci (index tie-break). Done when that choice is recorded.
5. Revise or stop
Stop on: all-pass; --track and score ≥ 0.8 with observed checks; --max-rounds exhausted (default: no revise); no score gain. Else revise only the swing winner on its weakest criteria and return to Step 3.
Done when a stop rule fires.
6. Report
criteria | generate | path | classify | ranking wi/ci | winner | rounds | stop | terminal. SUCCESS, ALLPASS, ALLFAIL, NO_IMPROVEMENT, BLOCKED.
Rationalizations
| Excuse | Reality |
|---|---|
| "I'll just do it myself" | Parent implements zero candidates. Fan out N. |
| "N should match difficulty" | N is --candidates. Do not invent it. |
| "Score them 1-10" | Discrete judge. Trust observed execution. |
Red flags
Parent wrote a candidate; same-tree or branch-as-parallel workers; N inferred; a 1–10; no swing wi/ci; no A/B swap.