Skip to content

Make RLCR completion state converge without no-op rounds #236

Description

@Harold-Finchwxx

Summary

An RLCR session completed all substantive work in its first delivery round, with strong evidence discipline, stop-rule compliance, and independent verification. However, completion-state friction caused additional no-change rounds that repeated validation and restated stable evidence before the loop exited.

The central issue is a state contradiction: a reviewer can confirm that every acceptance criterion is independently verified, no blocker or unexplained deferral exists, and a machine-checkable stop rule forbids further work, while still returning a continue/advance verdict.

Suggested improvements

1. Add acceptance-state convergence rules

When all acceptance criteria are independently verified, blockers are empty, deferrals are justified, and no legal work remains, the verdict should default to complete. A continue verdict without an explicit unmet criterion should be treated as inconsistent and automatically retried or corrected.

2. Add a completion-candidate path

Provide a dedicated path that does not require inventing a new mainline objective. It should only confirm acceptance coverage, unchanged work state, empty blockers, and satisfied stop conditions before entering final review.

3. Reuse verification for unchanged states

Persist a content fingerprint for the implementation, inputs, and evidence covered by validation. If the fingerprint and environment are unchanged, verify the fingerprint first and rerun full checks only when risk, freshness, environment changes, or contradictions justify it.

4. Structure freeze declarations

Separate these fields in summaries:

  • frozen semantic or decision content;
  • permitted append-only records;
  • functionality-equivalent implementation changes;
  • behavioral-equivalence evidence.

Tooling should prefill these fields from the change record and reject absolute claims that conflict with actual modifications.

5. Use incremental summaries and reviews

Keep stable facts in a machine-readable evidence ledger. Later rounds should report only changes since the previous round, newly rerun checks, intentionally reused checks, new findings, and the next state.

6. Decouple finding severity from round creation

Route findings into three levels:

  1. acceptance-blocking;
  2. requires closure but no implementation change;
  3. record-only follow-up.

Only the first level should force another implementation round. The second can use a post-review metadata step, and the third should remain queued.

7. Make stop rules first-class scheduler inputs

When a machine-checkable stop condition fires, prevent the scheduler from generating a new implementation objective and transition directly to evidence verification and exit. A reviewer requesting continuation must identify why the stop condition is invalid or insufficiently evidenced.

Expected benefits

These changes should eliminate no-op completion rounds, reduce repeated validation and narrative duplication, improve audit precision, and preserve strong independent review without weakening evidence or safety requirements.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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