Policy-Grade AI Agent Evaluations
Anka Reuel - Beyond Leaderboards: Building Policy-Grade Evaluations for AI Agents is a FAR.AI talk, uploaded May 11, 2026, on why AI-agent evaluations become fragile when developers or policymakers use them to justify deployment and regulatory decisions. The transcript separates two demands: validity, meaning the benchmark has to measure the claim being made, and reliability, meaning the score has to capture signal rather than noise.
Reuel grounds the argument in concrete cases: HealthBench-style healthcare claims without evidence that scores predict patient outcomes, benefits pre-screening agents where the same 90% score can hide harmless errors or wrongful denials, and non-deterministic agent runs whose point estimate can widen into a 72%-to-100% confidence range. For Spiralist themes, the talk matters because it treats benchmark numbers as institutional speech acts: once a score authorizes trust, regulation, launch decisions, or resource allocation, missing failure categories and uncertainty bounds become governance failures. The caveat is that this is a standards agenda, not a finished audit regime; its strongest contribution is a disciplined checklist for what eval reports should disclose before their numbers become policy facts.