OpenClinXR Slice Skill (Rubric Evaluator)

BLUF

After a work slice, produce a slice record and run the evaluator. Pass requires weighted checks against artifacts on disk. Scaffold TODOs, fabricated evidence paths, raised SIZE_FREEZE ceilings, and unfenced claim-safety language fail.

Commands (opt-in; not in default test / verify)

# Scaffold or re-evaluate a slice record
pnpm skill:slice:check -- --slice <slice-root> [--output <slice-record.md>] [--init-only] [--force-init] [--json]

# Direct evaluate
node --experimental-strip-types .agents/skills/openclinxr-slice/evaluator/evaluate-slice.ts \
  --slice <slice-root> --output <slice-record.md> [--json]

# Rubric regression suite (fixture corpus)
pnpm test:skill:slice
pnpm test:skill:slice:unit
pnpm test:skill:slice:integration

Required slice record sections

1. What changed — concrete packages/files/behaviors
2. Evidence passed — cite real paths under the slice tree
3. Remaining risk — residual risks; no clinical/exam claims
4. Validation performed — commands + outcomes + evidence paths

Template: templates/slice-record-template.md. Full scoring table: rubric.md.

Checks (weights / critical)

See rubric.md. Summary:

id	w	critical
required_slice_sections	3	yes
evidence_refs_exist	4	yes
touched_packages_tested	4	yes
no_guardrail_weakening	4	yes
claim_safety_language	3	yes

Pass: score ≥ 15 / 18 and no critical failure.

Fixture corpus

Fixture	Expected
passing-slice	pass
freeze-ceiling-raised	fail no_guardrail_weakening
fabricated-evidence	fail evidence_refs_exist
scaffold-boilerplate	fail sections + evidence + packages tested

AST note

Export and SIZE_FREEZE parsing uses typescript createSourceFile + node walk (evaluator/ts-ast.ts). Do not reintroduce regex export scanners.