What problem does it solve? Evaluating deliverables (code, designs, documents, slides, sites) by hand is inconsistent: the author grades their own work, reviewers get stale context, and pass/fail counts are eyeballed. This Skill provides a shared evaluation loop that picks the right method (tests, review skills, or independent judges), mechanically judges each round, and iterates fixes until the artifact reaches the pass line. ## Core Features & Use Cases - Three evaluation methods with selection rules: deterministic tests (A) first, existing review skills (B) second, and independent judge/reader subagents (C) only for subjective quality, always in A→B→C order. - Machine-enforced independence: a required provenance.json declares each evaluator's freshness, allowed inputs, and cache state; mismatches, missing evaluators, or side information stop judgment with exit 2. - Bounded unattended loop: run_round.sh chains judge_round.py (schema validation and threshold checks) and eval_state.py (cross-round state, convergence via blocking_count, max-rounds cap), producing a final report with a human gate outside the loop. - Use Case: After generating a landing page, declare the deterministic gates passed, spin up fresh judge and first-time-reader subagents with a rubric, and run rounds until scores clear the thresholds—then hand the final report to a human for the last decision. ## Quick Start Ask the AI to review your artifact with this skill and keep fixing and re-evaluating it until it passes the quality gate.
