Skip to content

Reviewer-prompt vs. byte-locked-plan harness-contract mismatch: 7 methodology improvements observed in a circuit-breaker exit #153

@Leopold-Fitz-AI

Description

@Leopold-Fitz-AI

Summary

A recent RLCR session exited via circuit-breaker stagnation at round 4 of 42 after 3 consecutive rounds of the reviewer re-raising the same structural finding against a byte-locked plan artifact that the harness physically forbids editing in-loop. The code-level loop worked well (rounds 0-3 closed real bugs). The structural-dispute loop did not: a contract mismatch between the reviewer's "deferred = incomplete" prompt and the harness's plan byte-lock rule produced 3 rounds of unactionable re-raise before the circuit breaker fired.

Fully sanitized analysis — no project-specific identifiers, paths, code, or domain terms.

7 suggestions

  1. Reviewer-prompt awareness of harness contracts: extend review-prompt schema with a "harness-locked artifacts" section and a "terminal-action verdict" the reviewer must use when the only close path is harness-external.

  2. Explicit exit-route signals: add a "terminal-dispute" verdict to both the implementer's update-request channel and the reviewer's verdict enum, so the loop can exit at round N+1 when both parties classify the dispute as structural.

  3. Pre-flight plan-prose audit: scan the tracked plan document at session start for placeholders that presuppose in-loop plan edits; fail-fast or stage a one-shot pre-loop amendment.

  4. Distinguish stagnation-as-failure from stagnation-as-playbook: when the implementer cites a canonical lesson in two consecutive rounds naming the expected exit-route, surface "playbook-execution" as a signal separate from "regression".

  5. Empirical-blocker artifact as first-class deliverable: add a dedicated attempt-and-revert round-contract template; require the rejection transcript verbatim; recognize it as a closing artifact for structural-dispute findings.

  6. Cap reviewer's authority to re-raise identical findings: after N consecutive rounds with the same evidence locator, force the reviewer to choose escalate / revise / accept rather than restate.

  7. Reviewer-side memory of prior implementer rebuttals: require the reviewer to explicitly mark each prior structured update-request as accepted / refuted (with new reasoning) / escalated, before raising the same finding again.

Context for prioritization

The core failure mode is two coexisting hard contracts that cannot both hold in the same in-loop step: the reviewer's prompt classifies "deferred = incomplete", while the harness classifies "the artifact under review is byte-locked at session-start". Any finding whose only close path edits the locked artifact is structurally unclosable in-loop. Suggestions 1, 2, 6, 7 directly add the missing vocabulary; 3 pre-empts the pattern at session start; 4 stops conflating playbook execution with regression; 5 turns the empirical-blocker side-product into a standardized deliverable.

Generated from a sanitized methodology-analysis report. Happy to share the full report (still sanitized) if useful for prioritization.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions