AANA Evaluation Pipeline

AANA attaches a verify-correct-gate stage after answer generation, separately measuring how well useful answers adhere to constraints.

• The flow consists of generator → verifier → corrector → gate → scorer, detecting factuality, safety, format, and uncertainty violations.
• In a comparison of 120 published cases, direct answers only reached a constraint check pass rate of 45.8%.
• The strongest AANA run recorded 98.3%, with the capability score rising from 0.662 → 0.922.
• The authors describe this as a research signal rather than a certified benchmark, and also provide documentation on architecture, evaluation design, and result interpretation, along with local sample execution examples.