AgentEval found 36 evidence faults that hash-only checks did not find
Independent checks agreed on all 30 complete episodes.
AgentEval keeps an assessment record with 3 parts: the actions of the agent, the evidence, and if the result is good at this time. In controlled changes to 48 evidence views, it showed 36 binding faults that hash-only checks did not find. Agent behavior did not change.
In controlled tests on 48 evidence views, AgentEval exposed 36 binding faults that hash-only checks missed, without changing agent behavior.
