mk:evaluate — Behavioral Active Verification

Step-file workflow that drives a running build, probes each rubric criterion via active verification, and produces a graded verdict with runtime evidence. Owned by the evaluator agent (Phase 3+).

When to Use

Activate when:

• User runs /mk:evaluate <target> with a URL, file path, or running-app handle
• A generator iteration completes and the harness needs a graded verdict
• After Phase 3 (build) and before Phase 5 (ship) for frontend/fullstack/CLI products
• When asked to "grade the running app", "check the build behaviorally", or "verify against the spec"

Skip when:

• The build has no runnable artifact (pure library, type-only package)
• The task is structural code review only — use mk:review instead
• The task is /mk:fix simple — overhead exceeds value

Hard Constraints

1. Active verification gate — every verdict MUST include non-empty evidence/ directory with at least one of: screenshot, HTTP response capture, CLI stdout+exit-code transcript. validate-verdict.sh rejects PASS verdicts with empty evidence and converts them to FAIL.
2. Skeptic persona enforced — load prompts/skeptic-persona.md at session start. Re-anchor before each criterion grading.
3. Max 15 criteria per session — split into multiple sessions if rubric composition exceeds. Heuristic: context overflow risk above this threshold.
4. No source code edits — evaluator owns tasks/reviews/*-evalverdict.md only. Never modifies source files.
5. Frontend default preset is pruned — frontend-app loads only product-depth, functionality, design-quality, originality (per Phase 2 v2.0.0 audit). Other 3 rubrics opt-in only.

Workflow

Execute via workflow.md. Step-file architecture — load one step at a time.

Step 1: Load Rubrics       → compose preset, load skeptic persona, parse sprint contract if exists
Step 2: Boot App           → start the build if not already running (skip if URL given)
Step 3: Probe Criteria     → drive browser/curl/CLI per criterion; capture evidence
Step 4: Grade and Verdict  → score each rubric; write verdict file; run validate-verdict.sh
Step 5: Generator Feedback → produce one-line fix guidance per FAIL/WARN; emit handoff message

Output

• tasks/reviews/YYMMDD-{slug}-evalverdict.md — structured verdict (YAML frontmatter + per-rubric sections)
• tasks/reviews/YYMMDD-{slug}-evalverdict-evidence/ — directory of screenshots, HTTP captures, CLI transcripts
• Returned to caller: PASS | WARN | FAIL + path to verdict file + generator feedback summary

Memory Write

After each completed evaluation, append a summary line to .claude/memory/review-patterns.md. Create the file with a header row if it does not exist:

| {date} | {artifact-id} | {verdict: PASS/WARN/FAIL} | {top-criterion} | {score} |

Use mkdir -p .claude/memory before the append to avoid silent failure on missing directory. This persists evaluation patterns across sessions for mk:elicit and mk:review to reference.

Verdict Schema

See step-04-grade-and-verdict.md for the canonical schema. Frontmatter fields: task, evaluator_run, rubric_preset, model, overall, weighted_score, hard_fail_triggered, iterations.

Gotchas

• Don't grade source code — that's mk:review's job. If you find yourself reading .tsx files instead of clicking buttons, you've drifted into static review
• Don't trust test pass claims — tests can pass against mocks while the real endpoint 500s. Run the build yourself
• Don't auto-load all 7 rubrics — frontend-app preset is pruned to 4 (Phase 2 v2.0.0). Loading the others duplicates work mk:review/security-rules/qa already do
• Don't issue PASS without evidence — the validator will reject it and convert to FAIL. Save yourself the round-trip
• Don't skip the skeptic persona reload — leniency drift is the dominant evaluator failure mode (Anthropic harness research)

References

Start

Read and follow workflow.md.