Source loop: pr-self-review (loops.elorm.xyz)
What the loop does
Review your own diff as a senior reviewer, fix findings, and repeat until you get N consecutive clean passes (default 3) with no critical findings — then open the PR. Convergence is defined by repeated agreement, because there is no binary command to check "is this code good".
Why valuable for MAP
MAP's Monitor / final-verifier run as single adversarial passes. For qualitative dimensions (correctness, edge cases, security smell) one pass can miss what a second clean pass would confirm. A "require N consecutive clean qualitative passes" convergence rule is a cheap confidence boost for high-risk subtasks where no deterministic oracle exists — and it generalizes the loop-until-dry / adversarial-quorum pattern MAP already uses elsewhere.
What MAP has today
- Monitor (single pass, binary
valid), final-verifier (single adversarial pass).
- No "converge to K consecutive clean passes" option for subjective gates.
Proposed scope
- Add an opt-in convergence mode to Monitor/self-review for risk-flagged subtasks: keep reviewing until K consecutive passes find nothing critical (K configurable, default 2).
- Cap total passes (bounded effort) and record each pass's findings for auditability.
- Keep deterministic gates single-pass; this applies only to qualitative review.
Notes
Distinct from N-vote adversarial verification (concurrent) — this is sequential-until-stable, which catches issues introduced by the previous pass's own fixes.
Part of #251
Source loop:
pr-self-review(loops.elorm.xyz)What the loop does
Review your own diff as a senior reviewer, fix findings, and repeat until you get N consecutive clean passes (default 3) with no critical findings — then open the PR. Convergence is defined by repeated agreement, because there is no binary command to check "is this code good".
Why valuable for MAP
MAP's Monitor /
final-verifierrun as single adversarial passes. For qualitative dimensions (correctness, edge cases, security smell) one pass can miss what a second clean pass would confirm. A "require N consecutive clean qualitative passes" convergence rule is a cheap confidence boost for high-risk subtasks where no deterministic oracle exists — and it generalizes the loop-until-dry / adversarial-quorum pattern MAP already uses elsewhere.What MAP has today
valid),final-verifier(single adversarial pass).Proposed scope
Notes
Distinct from N-vote adversarial verification (concurrent) — this is sequential-until-stable, which catches issues introduced by the previous pass's own fixes.
Part of #251