This repository publishes a bounded base-vs-AANA benchmark artifact on
PatronusAI/HaluBench.
The base path accepts every candidate answer as-is. The AANA path applies a
lightweight evidence-support gate over each answer and routes low-support answers
to revise.
This is not a trained hallucination classifier leaderboard submission. It is a
runtime-gate benchmark showing AANA's intended safety tradeoff: lower unsafe
acceptance of FAIL answers, with… See the full description on the dataset page:
https://huggingface.co/datasets/mindbomber/aana-halubench-grounded-gate.
