# Evolution 18 — release gates

| Gate | Status | Evidence | Note |
|---|---|---|---|
| A. Evolution 17 baseline reproduced | **PASS** | `baseline.md`; 154/155 before (one failure from another stopped agent's file in `cain/talos`, not E17), 316/316 after; E17 benchmark 20/20; v49 reproduces | |
| B. Constitutional invariants preserved | **PASS** (locally) | enforcer holds no approver private key (test + AB-04); no authorization → no execution (AD-02); no existing check weakened; 316 regression tests pass | nothing is wired to production, so this says nothing about live paths |
| C. Research and prior art for the selected capability | **PASS** with stated limits | `research_manifest.json` (21 sources, each fetched; 5 bad inherited citations removed) | no patent search |
| D. Canonical action-to-authorization binding tested | **PASS** | AB-01..10, CR-06, AD-03/04; mutants M01, M02, M06, M07, M10 killed | |
| E. Observer limitations documented and enforced | **PASS** | `observer_trust_model.md`; levels never exceed what sensors support (property tests; M11–M13 killed); ceiling in every grade | sensors are same host/kernel/uid |
| F. Causal evidence integrity tested | **PASS** | CO-01..05; graph tests; fixtures | |
| G. TOCTOU and concurrency tests executed | **PASS** | TT-01..09, CR-01..06, PV-03/04; M04, M16–M19 killed | `unlinkat` window and path executors remain (L-5, L-6) |
| H. Clean-room verifier regression preserved | **PASS** | `cain-interop-verify` (E17) and `cain-reality-verify` untouched; new verifier has no CAIN imports (test); stored fixtures verify | |
| I. Adversarial benchmark reproducible | **PASS** (same machine) | `python3 scripts/evolution18/run_evidence.py`; 82/82 executed | not reproduced on another machine or by another person |
| J. Performance overhead measured | **PASS** | `performance_results.md` | synthetic; ~19× slower than E17's unobserved path |
| K. Top-100 registry consistency | **PASS** | `top100_delta.md`; registry unchanged | |
| L. Invention register deduplicated | **PASS** | `ip_register_delta.md`; 3 candidates rejected as duplicates | |
| M. Valuation arithmetic reconciled | **PASS** | `valuation_reconciliation.md`; v49 reproduces, no arithmetic discrepancy; 4 open items recorded | tiering is a judgment, shown with alternatives |
| N. External review | **BLOCKED** | none performed | needs an independent reviewer |
| O. Production readiness | **BLOCKED** | not wired to any live path; single host; sensors not separated from the enforcer's operator | |