Bidirectional NLI scores source→summary and summary→source, takes min.
Baseline calibration (0.20) shifts expected NLI noise floor to zero.
All three task types now production-grade: QA 3-4%, dialogue 4.5%,
summarization 10.5%. 2051 tests, 0 failures.

Co-Authored-By: Arcane Sapience <protoscience@anulum.li>