Evals and Safety: From Final Scores to System Reliability
Evaluation is broadening to execution paths, evidence grounding, omissions, calibration, artifact consistency, stopping behavior, and independent checking. Guidance that agents should produce inspectable disagreement and halt before acting aligns with this shift, while verifier-guided OOD generation such as ActFlow remains promising but insufficiently evidenced. LLM-judge reliability and uncertain ground truth remain central constraints.
Sources (7)
Updated Sep 16, 2026