Repository navigation
Research convergence-based review termination versus fixed correction caps #132
Description
Goal
Determine whether bounded review should continue using fixed correction caps or adopt a measurable convergence-based stopping rule after the 1.0 trust contract is stable.
Research questions
- Which observable signals distinguish genuine convergence from reviewers merely repeating or rephrasing findings?
- Can a convergence rule terminate faster without accepting unresolved high-severity defects?
- How should reviewer disagreement, infrastructure failure, malformed output, and candidate changes reset or stop the process?
- What hard ceiling is still required so convergence logic can never create an unbounded review loop?
- Does the answer differ for incremental checkpoints versus the final whole-candidate gate?
Method
- Use completed review receipts and the post-1.0 cadence dataset; do not change the 1.0 gate while gathering evidence.
- Define candidate convergence metrics before scoring outcomes.
- Replay representative clean agreements, verdict splits, repeated findings, newly discovered blockers, and infrastructure failures.
- Compare fixed-cap and convergence policies on quality, review rounds, elapsed time, quota, false certification, and human escalation.
- Recommend retain, experiment, or adopt with explicit confidence and rollback criteria.
Acceptance criteria
- Dataset and scoring method are reproducible and privacy-safe.
- Fixed-cap behavior is the control.
- High-severity unresolved findings can never be classified as convergence.
- Every proposed policy has an absolute round/time ceiling and fail-closed behavior.
- The recommendation separates incremental checkpoints from final delivery review.
- Any implementation is split into a later issue with TDD, proof, and independent review.
Scope
Research-only work in the 1.1.0 milestone after 1.0 ships. It does not block 1.0 and must not change current review defaults.
Activity
BaseInfinity commentedon Aug 23, 2026
Consumer control result: the fixed correction cap prevented recursive review hell
An ~18-hour
BaseInfinity/m180-jumpseatsequence completed three bounded issues and their lifecycle reconciliation without an unbounded review loop:The suite grew from 2,472 to 2,491 passing tests. Seven implementation/lifecycle/correction PRs merged. No reviewer was asked to keep searching after a clean terminal review, and no unchanged-input review was used to wish findings away.
What appears to have worked:
Recommendation for this research issue: retain the fixed cap as the control. Test convergence only as a way to stop earlier, never as permission to exceed the absolute cap or accept unresolved P0/P1 findings. The useful convergence signal here was exact-finding closure followed by one clean frozen-candidate review—not repeated reviewer agreement rounds.