# Confidence and Uncertainty

Confidence is a bounded `0…1` statement about the support for a contextual estimate. It is not certainty about identity and not clinical validity.

Every state and assessment MUST expose:

- score and confidence;
- evidence count and effective weight;
- context coverage;
- agreement/contradiction;
- oldest/newest evidence and half-life;
- missing-data and calibration warnings;
- method and method ceiling.

## Default gates

| Confidence | Permitted default |
|---|---|
| `< 0.40` | Observe, disclose uncertainty, or ask; no drift intervention |
| `0.40–0.64` | Suggest collecting evidence; optional low-stakes question |
| `0.65–0.79` | Explain/recommend; reversible low-impact Ally modulation if policy explicitly permits |
| `≥ 0.80` | Zone-authorized AI/Actor correction or escalation, still subject to evidence counts and Authority |

A deterministic hard-boundary event (for example, an Action requests an unauthorized Connection) may be constrained without a behavioral confidence estimate because the policy predicate itself is proven. A probabilistic boundary inference requires the boundary's confidence gate.

## Missing data

No observation means `unknown`, not zero. Expired evidence is excluded. If a dimension lacks enough contexts, the engine emits `insufficient-context-coverage`. If equal-quality evidence supports incompatible readings, it emits `contradictory-evidence` and lowers confidence.

## Calibration

Predicted confidence MUST be compared with adjudicated outcomes by dimension, entity type, context, language/culture where lawful, and response consequence. Overconfident dimensions are capped or suspended. Calibration data MUST NOT be repurposed into human ranking.