Skill Eval

Evaluate whether one skill improves the produced work. Do not award credit for headings, claimed procedure, or a report that merely looks complete.

Input

Require:

• exactly one target: <INDEX name> or subject_path: <path>;
• phase: calibrate | verify;
• one surface with adapter: codex | claude-code | cursor | opencode | cline and an exact model name;
• evals/evals.json and a resolvable old-skill or no-skill baseline.

Hash the complete subject package with scripts/hash-skill-package.mjs. Every terminal result belongs only to that hash. If the package changes, re-run the evaluation.