This repository publishes a bounded base-vs-AANA benchmark artifact on
wandb/RAGTruth-processed.
The base path accepts existing model outputs as-is. The AANA path applies a
lightweight evidence-support gate over each model output and routes low-support
outputs 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… See the full description on the dataset page:
https://huggingface.co/datasets/mindbomber/aana-ragtruth-grounded-gate.
