This repository publishes a bounded calibration artifact for AANA grounded
hallucination and unsupported-claim gates.
The calibration goal is to reduce false positives while preserving high recall
on hallucinated, inaccurate, or unsupported outputs. This is a threshold
calibration pass over existing deterministic evidence gates, not a trained
classifier or leaderboard submission.
