Speck Self Eval

Speck self-eval (A1-lite)

Runs seeded artifacts through the candidate-corpus gate anchors and deterministic defect evaluator. Use before changing named Speck gate capabilities. This is a contract smoke test, not runtime agent-behavior proof.

Framework-repository-only: if tests/eval/score.sh or its immutable baseline is absent, STOP. Do not fabricate or copy a scorecard into a product repository.

Do

1. Run bash tests/eval/score.sh --check while iterating; it must exit nonzero on a harness error, wrong fixture verdict, or regression from reports/baseline.json.
2. Read stdout + tests/eval/reports/latest.md.
3. Run bash tests/eval/score.test.sh; require the corpus-deletion and wrong-verdict mutations to turn red.
4. After the candidate is final, run bash tests/eval/score.sh once to refresh latest.md. Never rewrite baseline.json as a side effect of scoring.

Output

Fail-closed scorecard with per-fixture verdicts, harness-error count, and immutable-baseline regressions.