8 of 48 agent sessions had hallucinated refund timelines, caught by jev judge

strickvl · x · 2026-09-24

To test whether TypeSafe's jev judge was just number-matching, strickvl edited one field at a time in a session: a '30 days' from the policy tool passes, but promising a customer 30 days fails — even though a 30 appears in the tool results (the return window, not the refund time), showing real semantic understanding.

In practice, jev gave the timeline question p(yes)=0.93; across 48 sessions imported from their tracing provider, 8 contained hallucinated refund timelines, easy to confirm manually once surfaced.

Related event: Testing the jev Judge: 8 Hallucinated Refund Timelines Found in 48 Sessions(2 posts)→

Original post →

More from coding & agent

coding & agent channel →