Improvement Loop

Improvement is a bounded behavior-improvement loop.

Map keys: promise.improvement, rule.evidence-gaps, rule.cost-and-proof-freshness, rule.reviewable-artifacts. Evidence path: deterministic packet tests plus a live held-out improve cycle. Evidence status: satisfied on the dev/skill surface. Next action: extend the live held-out improve proof beyond the dev/skill orientation surface as new improvement targets are adopted. Terms covered here: improve prepare-input, search budget, checkpoint, frontier, proposal, held-out validation, protected checks, runtime fingerprint, blocked readiness, reuse.

Maintainer Promise

Improvement starts from an explicit behavior target and budget, and records what changed, which checks were protected, which results were reused, and which held-out checks still guard regressions.

Subclaims

  • improve prepare-input records the target claim, budget, and protected checks before the search loop runs.
  • improve search run and revision-artifact assembly preserve checkpoint state and frontier promotions so the loop is resumable.
  • Blocked-readiness conditions surface through the improve packet rather than being hidden behind repeated retries.
  • Held-out evaluation guards regressions; an improvement claim is not treated as satisfied without a held-out proof.

Evidence

Evidence Gaps

  • The live held-out improve proof currently covers the dev/skill orientation prompt target. Owner: maintainer. Next action: extend the live held-out improve proof to additional improvement targets as they are adopted.