Skip to main content

GOV 13 REVIEW LOOPS AND COMPLETION DISCIPLINE

A glass puck reaches the closure seam of a circular metal inspection track.

This page embeds the canonical rule text and adds commentary after each section to explain why the section exists.

Governance: Review Loops and Completion Discipline​

Implemented is not complete.

Governed agent work should close the loop through review, evidence, and feedback incorporation before claiming completion.

Commentary: Separates first-pass output from trustworthy completion.

Core Principle​

  • GOV-13-REV-001 A first-pass implementation is an intermediate state, not a trustworthy completion signal.
  • GOV-13-REV-002 Completion requires explicit review of the produced change and its evidence, not only confidence in the act of producing it.
  • GOV-13-REV-003 Review loops should be designed to strengthen correctness, maintainability, and harness learning.

Commentary: Defines review as a required control loop, not optional polish.

Required Review Loop​

  • GOV-13-REV-004 Before claiming a governed work unit complete, the agent should perform a targeted self-review against intent, scope, diff quality, evidence quality, and residual risk.
  • GOV-13-REV-005 The review should ask at minimum:
    • does the change satisfy the intended requirement or issue,
    • does the diff stay inside scope,
    • is the evidence direct enough,
    • are docs/spec/traceability updated where required,
    • what remains unverified, blocked, or deferred.
  • GOV-13-REV-006 If the review reveals material weakness, the agent should continue the loop rather than claiming done and hoping later review catches it.

Commentary: Makes internal review explicit and prevents premature completion claims.

Feedback Incorporation​

  • GOV-13-REV-007 Human, automated, or agent review feedback should be incorporated as part of the same governed work loop until the active review scope is honestly resolved, re-scoped, or blocked.
  • GOV-13-REV-008 Responding to feedback should include updating affected evidence, docs, and traceability when those artifacts are impacted.
  • GOV-13-REV-009 Repeated feedback themes should be considered candidates for governance or harness promotion.

Commentary: Turns review comments into an execution loop and a governance learning loop.

Feedback Assimilation and Self-Improvement​

  • GOV-13-REV-010 When approved or edited feedback reveals a reusable lesson, the agent should compare the original draft with the approved or edited version and identify the concrete pattern of change.
  • GOV-13-REV-011 That pattern should be translated into a short future rule, drafting rule, governance rule, or harness adjustment when the lesson is durable enough to matter again.
  • GOV-13-REV-012 Material feedback lessons should be persisted to the relevant governed artifact before long posting phases, long continuation phases, or likely context-loss transitions when practical.
  • GOV-13-REV-013 New learned rules should be checked against existing rules for contradiction, overlap, or duplication and then merged or reconciled instead of appended blindly.

Commentary: Makes review-driven learning explicit so approved edits improve the future system, not just the current artifact.

Completion Semantics​

  • GOV-13-REV-014 Completion claims should distinguish clearly between implemented, verified, reviewed, and released states rather than collapsing them into one vague "done".
  • GOV-13-REV-015 A governed work unit is not complete if review, evidence, or follow-up handling is still materially open.
  • GOV-13-REV-016 If confidence is partial, the claimed state should say so explicitly, for example blocked, partial, scoped-complete, or awaiting review.

Commentary: Improves handoff clarity and reduces false certainty in progress reporting.

Escalation and Stop Conditions​

  • GOV-13-REV-017 When repeated review/fix loops stop producing clear progress, the agent should escalate the blocker, ambiguity, or missing capability instead of performing aimless churn.
  • GOV-13-REV-018 When judgment is required beyond the current governance and evidence, the agent should make the decision boundary visible instead of silently guessing.
  • GOV-13-REV-019 Review/fix progress must be judged by evidence movement, not mutation count, token use, lines changed, commits, or confident status prose.
  • GOV-13-REV-020 Equivalent failures should be classified by stable failure signature so superficial message changes do not reset the circuit-breaker count.
  • GOV-13-REV-021 A repeated attempt counts as progress only when it narrows the hypothesis, adds diagnostic evidence, restores a regressed gate, or improves direct verification against the governed claim.
  • GOV-13-REV-022 When a circuit breaker triggers, the current work unit must enter scoped blocking and produce the incident evidence required by gov-02 before another implementation attempt begins.
  • GOV-13-REV-023 A fresh session, worker, or prompt does not itself satisfy a resume condition; materially new evidence or constraints must change the recovery path.

Commentary: Prevents endless churn loops and forces explicit escalation boundaries.

Anti-Patterns​

Avoid these failure modes:

  • equating implementation with completion
  • claiming done immediately after the first passing check
  • leaving review comments or feedback themes unincorporated without explicit state change
  • treating approved edits as one-off fixes with no learning loop when they reveal a reusable pattern
  • adding duplicate learned rules without contradiction checking
  • resetting a failed-loop counter because the same failure was reworded
  • treating more code, commits, tokens, or tests as proof of progress
  • restarting the same blocked approach in a fresh session without new evidence
  • using a summary as a substitute for a real review loop
  • moving on while material review debt remains hidden

Commentary: Names the output-first behaviors this rule is meant to correct.