How do I tell whether an eval case has gone stale?

An eval case has gone stale when the behavior it expects no longer matches what the agent is supposed to do, so a pass or a fail on it says nothing about the agent. A case like that tests agreement with its own past.

The signal is the capture date read against what has changed since. A case goes stale once the prompt, tool, or model it was written against has moved and nobody has re-verified that the expected behavior still holds. It still runs and still produces a score, which is what makes staleness dangerous: nothing errors, the number just quietly stops meaning what it used to.

You can’t catch this without a capture date on the case in the first place. In practice that means re-pulling cases against current traffic periodically, rather than trusting a set once and leaving it alone, and re-checking the golden set whenever the agent or the judge model moves.