# CAIN ORACLE — World-Model Assurance & Calibration

`cain/oracle/worldmodel.py`, `cain/oracle/calibration.py`.

## WorldModelAssuranceEngine
Deterministic rules over historical accuracy, calibration error, horizon, environment similarity, distribution shift,
staleness, independent evidence, disagreement, recent failures, causal validity, intervention coverage, unseen-state
exposure and environment binding → state:

`TRUSTED · CONDITIONALLY_TRUSTED · DEGRADED · STALE · SHIFTED · UNVERIFIED · FAILED · QUARANTINED`.

Each state has an **authority ceiling** (TRUSTED 1.0 → UNVERIFIED/FAILED 0.0). **World-model authority never exceeds
world-model evidence.** `may_authorize(state, consequence_class)` requires a stronger state for a higher consequence
(P177, P178).

## Prediction calibration
`CalibrationTracker` records predicted vs observed consequence classes and confidence; `summary()` reports mean error,
false-confidence rate, bias and calibration state (`CALIBRATED · UNDERCONFIDENT · OVERCONFIDENT · UNCALIBRATED ·
UNKNOWN`).

## Prediction error → authority impact
`PredictionAuthorityImpactEngine.assess` — a prediction error must **reduce future authority**, not merely be logged.
A **confident wrong** prediction (confidence ≥ 0.8) is weighted higher than an uncertain one (P179, P180); a
conservative over-prediction does not reduce authority (P181). Actions: `REDUCE_AUTHORITY`, `RESTRICT_ACTION_CLASS`,
`ESCALATE_EVIDENCE`, `RECALIBRATE_MODEL`, `RECERTIFY`, `PENALIZE_OVERCONFIDENCE`; targets: agent/model/world_model/
action_class/consequence_class.

## Status
IMPLEMENTED, TESTED.
