Skip to content

v0.6.2 — Proportionate Agent Verification

Latest

Choose a tag to compare

@oneingan oneingan released this 19 Oct 11:13

Changed

  • Made the narrowest sufficient verification the default, widening for required obligations, uncovered risks, or uncertain impact without weakening safety, integration, CI, release gates, or accountable human acceptance.
  • Clarified evidence validity and reuse, conditional additional verification, and stopping once acceptance is supported instead of automatically duplicating proof.
  • Bounded stalled work and recovery by outcomes rather than repeated activity, without treating exhausted effort as success.
  • Added routine-verification routes from AGENTS.md, the main context map, and existing retrieval metadata; refined existing playbook, orchestration, and review guidance without a new canonical layer or checklist items.

Evaluation

  • Reviewed long-form articles and vendor documentation, excluding standalone tweets; model-specific guidance, evaluation tooling, and raw artifacts remain outside the vendored core.
  • Preserved the initial eight-run decision-only pilot and its disclosed global-prompt contamination; it did not demonstrate an efficiency gain.
  • Ran six executable synthetic-consumer sessions with identical captured effective prompts/tools and explicit isolation: four completed tasks and two correctly blocked mandatory-integration handoffs, with no observed scope or acceptance regression.
  • The candidate consulted the verification policy in 3/3 runs versus 0/3 at baseline. Candidate elapsed times were higher in these observations; no speedup or human-review saving is claimed.
  • No application-wide suites or same-check/same-input reruns were executed. An additional pre-fix reproduction supplied useful evidence rather than redundant verification.

Validation

  • Passed nix shell nixpkgs#yq nixpkgs#ripgrep nixpkgs#python3 -c ./tooling/run-checks.sh.
  • Passed git diff --check.

Evidence and limits

Full changelog: v0.6.1...v0.6.2