three-stage-eval

What problem does it solve?

Ensures implemented artifacts meet their seed specifications by providing a repeatable, gated evaluation pipeline that catches mechanical issues, semantic mismatches, and unresolved high-stakes disagreements before merge or deployment.

Core Features & Use Cases

• Mechanical Verification: Run syntax, lint, test, and coverage checks to catch immediate build and quality failures.
• Semantic Evaluation: Map artifact outputs to acceptance criteria, gather evidence, compute AC compliance and an overall score, and quantify drift.
• Consensus Review: Trigger a multi-perspective deliberation (Proposer, Devil's Advocate, Synthesizer) when scores are borderline or drift is high to reach a 2/3 decision.
• Use Case: Pre-merge gate for libraries, infrastructure configs, or agent skills to prevent regressions and ensure alignment with spec-first seeds.

Quick Start

Use the three-stage-eval skill to evaluate /path/to/implementation against /path/to/seed.yaml and save the resulting evaluation report to the workspace.