# CAIN ORACLE — Test Report

| Suite | Command | Result |
|---|---|---|
| ORACLE tests | `pytest tests/test_oracle.py` | **15/15 passed** |
| Formal properties (v1) | `run_property_suite11()` | **15/15 passed (P176–P190)** |
| Formal properties (v2) | `run_property_suite_oracle2()` | **20/20 passed (P191–P210)** |
| Adversarial lab | `run_oracle_lab(cases=100000)` | **99,974 executed / 99,974 blocked / 0 violations**, 37 categories |
| Mutations | `run_oracle_mutations()` | **24/24 killed** |
| Predictive chaos | `OracleChaosEngine.run()` | **17/17 faults fail-closed** |
| Clean-room verifier | `python3 cain-oracle-verify/verify.py prediction\|certificate\|failure` | valid accepted; tampered rejected |
| Latency benchmark | `python3 scripts/cain42_evolution/bench_oracle.py` | per-op p50/p95/p99 in `artifacts/oracle/ORACLE_BENCH.json` |
| Live wiring | `install_hosted_pipeline()` | **default_stages_v19 installed by default (includes 4 ORACLE stages)** |

## What is exercised
prediction integrity/binding/expiry/nonce/model/environment; world-model assurance states + authority ceilings +
`may_authorize`; calibration overconfidence; prediction-error → authority impact; counterfactual robustness;
simulator admissibility; reality override + untrustworthy-observation; certificate semantics (world model, simulator,
counterfactuals, evidence independence, consequence bound, expiry, action, environment, nonce); **multi-oracle
disagreement (independent vs correlated agreement, severe contradiction, blind spots); evidence-independence graph
(shared dataset/simulator/sensor roots); prediction-failure learning governance (no silent authority increase;
approval required; increases rejected; foreseeable weighting); governed failure records + clean-room verification;
predictive chaos fail-closed; FAST/STANDARD/HIGH_ASSURANCE path selection**; property suites; lab; mutations; hosted
stages (installed by default); clean-room verifier.

## Limits
Same-project tests only; no third-party validation; prediction reliability is a heuristic, not a guarantee; no real
physical system; the lab and mutations are self-defined by the project; the benchmark is a local in-process
micro-benchmark, not a production SLA.
