Add 5 scenario playbooks

docs/evidence/PLAYBOOK_BENCHMARK_REGRESSION.md

@@ -0,0 +1,13 @@
+# Scenario Playbook: Benchmark Regression
+
+**Scenario:** A benchmark AUROC drops by more than 0.02.
+
+## Steps
+
+1. Verify the regression is real (not a fluke) by running the benchmark twice.
+2. Check what changed since the last passing run (`git log --oneline`).
+3. Isolate the change that caused the regression using `git bisect` if needed.
+4. If the regression is caused by a deliberate tradeoff (e.g., safety penalty increase), document the tradeoff.
+5. If the regression is unintended, fix or revert the change.
+6. Update `docs/evidence/METRICS_CURRENT.md` with the new value.
+7. Run `make bench-gate` to verify the fix.

docs/evidence/PLAYBOOK_CHANGED_OUTPUT.md

@@ -0,0 +1,16 @@
+# Scenario Playbook: Changed Command Output
+
+**Scenario:** A CLI command output format changes, potentially breaking downstream consumers.
+
+## Steps
+
+1. Determine if the change is intentional or accidental.
+2. If intentional:
+   - Update all documentation that references the old output format.
+   - Update any tests that check the output format.
+   - Add a changelog entry describing the change.
+   - Consider backward compatibility or a migration path.
+3. If accidental:
+   - Fix the code to restore the expected output format.
+   - Add a test to prevent future regressions.
+4. Run the full test suite to verify no other code depends on the old format.

docs/evidence/PLAYBOOK_OVERCLAIM_REVIEW.md

@@ -0,0 +1,14 @@
+# Scenario Playbook: Overclaim Review
+
+**Scenario:** A reviewer flags overclaiming language in a PR.
+
+## Steps
+
+1. Run `python scripts/safety/check_claims.py` to see all flagged phrases.
+2. For each flagged phrase, determine if it's a genuine overclaim or an allowed exception.
+3. Rewrite overclaiming phrases using the project's allowed terms:
+   - "computationally nominated candidate" (not "drug candidate")
+   - "predicted antimicrobial peptide" (not "proven antimicrobial")
+   - "selected by reproducible pipeline" (not "validated by computation")
+4. Re-run the claim scanner to verify the fix.
+5. If the original author disagrees with a flag, document the reasoning.

docs/evidence/PLAYBOOK_PROOF_CONFUSION.md

@@ -0,0 +1,11 @@
+# Scenario Playbook: Proof Confusion
+
+**Scenario:** Someone cites a computational score as biological proof.
+
+## Steps
+
+1. Refer them to `docs/evidence/PROOF_LADDER.md` — levels 0-4 are computational.
+2. Explain that only levels 5 (lab assay) and 6 (independent replication) constitute biological evidence.
+3. If the confusion is in project documentation, update the document to clarify the proof level.
+4. If the confusion is in an external communication, provide a correction.
+5. Log the incident in the decision log as a documentation clarification.

docs/evidence/PLAYBOOK_STALE_ARTIFACT.md

@@ -0,0 +1,17 @@
+# Scenario Playbook: Stale Artifact Finding
+
+**Scenario:** An artifact (schema, benchmark output, config) is stale and doesn't match the current code.
+
+## Steps
+
+1. Identify what's stale: check the artifact's schema version vs current code.
+2. If the artifact is a benchmark output, regenerate it: `make bench-500`.
+3. If the artifact is a schema, update it to match the current code.
+4. If the artifact is a config file, update it or document why the discrepancy is acceptable.
+5. Update any documentation that references the stale artifact.
+6. Add a test to catch similar staleness in the future.
+
+## Prevention
+
+- Run `make regenerate-all` before major releases.
+- CI should validate schema versions against code.

tests/test_playbooks.py

@@ -0,0 +1,21 @@
+"""Tests for playbooks."""
+from pathlib import Path
+
+PLAYBOOKS = [
+    "PLAYBOOK_OVERCLAIM_REVIEW.md",
+    "PLAYBOOK_PROOF_CONFUSION.md",
+    "PLAYBOOK_BENCHMARK_REGRESSION.md",
+    "PLAYBOOK_STALE_ARTIFACT.md",
+    "PLAYBOOK_CHANGED_OUTPUT.md",
+]
+
+
+def test_all_exist():
+    for p in PLAYBOOKS:
+        assert Path(f"docs/evidence/{p}").exists(), f"Missing: {p}"
+
+
+def test_each_has_steps():
+    for p in PLAYBOOKS:
+        text = Path(f"docs/evidence/{p}").read_text()
+        assert "Steps" in text or "steps" in text