ADR-0003: Transparent rule-based confidence model

Status: Accepted

Context

Confidence is the core differentiator, and it must be auditable. The spec says "ACH/confidence: transparent scoring (rules)".

Decision

• base = METHOD_PRIOR[method] x RELIABILITY_WEIGHT[source grade]. The method priors are observed 0.9 > inferred 0.7 > stated 0.6 > predicted 0.5, and the source grade uses the A-F scale.
• Corroboration is a noisy-OR over independent sources: only the best claim per source counts, so repeated reports from one feed do not inflate confidence.
• Contradiction discounts the result by 1 - 0.6 x strongest contradicting claim. ATTRIBUTED_TO is exclusive per subject.
• The result is capped at 0.99. Nothing is certain.
• explain() returns every input. brier() is provided for calibration benchmarking.

Consequences

• The model is simple, explainable and deterministic. Its constants are uncalibrated guesses (a Grade D TODO: calibrate them against labeled range data using Brier scores).
• Noisy-OR assumes sources are independent, but real feeds copy each other. Source-lineage deduplication is a TODO.