Premise Check

Core principle

Kill wrong premises before they become plans. The output is a survival list, not a score.

When this, not scored-iteration

Never score. Never iterate. One round, then hand survivors to superpowers:writing-plans.

The three non-negotiables

1. At least one reviewer gets primary-source access — the raw data, repo, transcripts, or original documents. A reviewer given only your summary produces inflation, not findings. If no reviewer can be given access, mark that reviewer 증거없음 and give its verdict zero weight.
2. Every objection ships something a third party can check without asking the reviewer — a reproduction command, or an exact source locator (file:line, URL) with the verbatim passage. 이 명령을 돌리면 이 수치가 나온다 and 이 출처 이 줄은 반대를 말한다 are both valid. "우려된다" / "리스크가 있다" is rejected without appeal.
3. Output is a three-way classification, never a total. 유지 / 근거붕괴 / 부작용초과. A composite score across independent proposals is meaningless — forbid it.

Classification rules (mechanical — apply in order)

1. 근거붕괴 — a necessary premise was falsified, or the proposal's own mechanical completion criterion is now invalid. If the metric that would prove the proposal done has been refuted, the proposal has no acceptance test and dies here regardless of how appealing it still sounds.
2. 부작용초과 — the side effect is irreversible (data exposure, git history, published state, destroyed evidence) and the reviewer names the concrete mechanism and the state that cannot be restored. Severity labels do not qualify a proposal for this class; a named irreversible mechanism does. Both conditions, or it does not qualify.
3. 유지 — everything else, including "evidence weakened", "mechanism unproven", and "no local data yet". Weakened is not dead. Attach the condition that would re-establish the evidence, and hand it to planning with that condition attached.

A proposal you want to drop but that lands in 유지 means you are carrying an unstated reason. State it as a falsifiable claim or keep the proposal.

Run sequence

1. Brief — raw data only. Strip your conclusions, priority order, and recommended ranking. Shuffle proposal IDs so ordering carries no signal. Write it to a file; do not paste your reasoning.
2. Dispatch ≥2 reviewers in parallel. Give each a distinct lens — one that verifies claims against primary sources, one that attacks the premise and predicts new failure modes. Means are free: Agent tool (Claude subagent) or codex exec --sandbox read-only via Bash. Mixing model families is recommended, not required — evidence supports primary-source access, not cross-family review.
3. Collect. Reject objections that ship neither a reproduction command nor an exact source locator with the verbatim passage (non-negotiable 2).
4. Drop severities that carry no variance. A reviewer told to break things will break things. When a reviewer assigns essentially one severity across its items, that label ranks nothing — discard its severities entirely and judge each item on its evidence alone. Do not downgrade: on a two-level scale a downgrade annihilates every finding, including the correct ones.
5. Verify what changes your conclusion. Re-run any measurement a reviewer says is wrong. Do not take the refutation on trust either.
6. Classify and stop. Survivors go to planning. Everything else is dropped, with the falsifying evidence recorded.

Re-judgment rule

New facts trigger re-judgment. The same evidence never does.

Re-scoring the same material is how a bounded check turns into an unbounded loop. This is not hypothetical: an attempt to fold this premise gate into scored-iteration as one more scoring axis failed to converge across three rounds and was abandoned. The two skills stay separate for that reason — a premise is falsified or it is not, and that verdict does not improve by iterating on it.

Red flags — stop

• A composite score, a quality percentage, or an "N/100" appears in a reviewer verdict or in the classification output. (Counting findings is not scoring.)
• A second round opens on the same evidence.
• Every reviewer saw only your summary.
• An objection survives with neither a reproduction command nor an exact source locator.
• You accept a refutation of your own measurement without re-running it.
• The brief carries your recommended order, priority labels, or rationale.
• Survivors go straight to implementation instead of through planning.