# CAIN REALITY — Test Report

| Suite | Command | Result |
|---|---|---|
| REALITY tests | `pytest tests/test_reality.py` | **10/10 passed** |
| Formal properties | `run_property_suite_reality()` | **30/30 passed (P241–P270)** |
| Adversarial lab | `run_reality_lab(cases=100000)` | **99,995 executed / 99,995 blocked / false-verified-outcome = 0**, 35 categories |
| Mutations | `run_reality_mutations()` | **20/20 killed** |
| Reality chaos | `RealityChaosEngine.run()` | **16/16 faults fail-closed** |
| Clean-room verifier | `python3 cain-reality-verify/verify.py snapshot\|expected\|certificate\|bundle` | valid accepted; tampered rejected; bundle recomputes |
| Latency benchmark | `python3 scripts/cain42_evolution/bench_reality.py` | per-op p50/p95/p99 in `artifacts/reality/REALITY_BENCH.json` |
| **REAL external canary** | `python3 scripts/cain42_evolution/reality_canary.py` | **COMPLETED**: real effect → real observation → MATCH → CAUSALLY_SUPPORTED → VERIFIED_GROUND_TRUTH → clean-room valid |
| Live wiring | `install_hosted_pipeline()` | **default_stages_v21 installed by default (includes 3 reality stages)** |

## The decisive result (section 52)
A real message was published to a real external system (ntfy.sh), observed independently, reconciled to MATCH, causally
attributed via a governed token, sealed in a `GroundTruthCertificate`, and reproduced by the clean-room verifier
(`clean_room_certificate_valid=true`, `clean_room_bundle_reproduced=true`). No simulation was used in that path.

## What is exercised
snapshot/expected integrity + agent-source rejection; reconciliation states (MATCH/PARTIAL/MISMATCH/CONTRADICTED/
UNKNOWN/STALE/UNAVAILABLE); field rules; contradiction; multi-observer independence; causal attribution
(supported/correlated/ambiguous/contradicted); concurrent actor; temporal validity + policy freshness; eventual
consistency; drift; world-model calibration; ground-truth semantics + tamper/replay/expiry; deterministic replay;
git-external observation; property suite; lab; mutations; chaos; hosted stages.

## Limits
Same-project tests only; no third-party validation; one real external system demonstrated; physical AI SIMULATED; the
lab, mutations and benchmark are self-defined / local.
