11 — Safety / No-Fake Audit (Workstream J)

invariant	result	evidence
No production DOT executed	✅	KG pair usage_count sum=0, last_executed=NULL; no dispatcher exists
No REAL_RUN	✅	0 REAL_RUN rows; real_run_enabled=false; observe fn refuses REAL_RUN
No fake verified	✅	verified_candidates_v3 = only job:cut; dot:kg = runner_contract_missing
No process birth / canon	✅	birth_registry 1,158,319 before==after; 0 process born
No approval approved	✅	no approval_requests touched
No event activation	✅	7 process.* still active=false; 0 event_outbox emitted
No IU / document edit	✅	only KB report authoring (disclosed provenance births)
No workflow / routing run	✅	none invoked
Idempotency	✅	re-call of runner → same process_run_id, 0 new rows
Rollback / cleanup exists	✅	v4_rollback.sql (DROP 4 views + fn); DELETE clause for runner obs

DDL guard / trigger surface

trigger_guard_alerts = 129 unchanged. New objects = 1 function + 4 views + 1 observation row (via existing fail-closed functions). No business-table schema change; no new birth trigger attached (new function/views carry none).

Honesty notes

• The only evidence_type written is SIMULATED_DRY_RUN — correct, because no logic was executed.
• DRY_RUN and REAL_RUN remain at 0 rows system-wide.
• The "true dry-run capability exists for job:cut" claim is backed by the iu-cutter o7 sidecar report + 366/366 tests, not by a re-run this session.

KB provenance births (disclosed)

Publishing the ~17 report/checkpoint/fixture docs creates knowledge_documents provenance births (species ai_support) via the normal KB→Directus path — document births, not process/taxonomy/approval births. This is the inherent, expected cost of publishing, consistent with prior macros. PG-schema births from the runtime work = 0.

Audit: PASS.