Hearsay evaluates whether agent-run records can support later audits without requiring trust in the harness that produced them. It audits sixteen frameworks, then has three blinded LLM examiners and a human panel inspect 140 runs from five harnesses across fourteen tasks while mechanically checking who wrote every cited excerpt.
Examiners identify the correct fault in 74 to 91% of runs, but fewer than one citation in ten refers to material not written by the harness. An external append-only log, reconciled against the harness record in both directions, detects all 28 omissions and fabrications introduced during execution; a hash chain over the harness record detects none because it authenticates whatever the harness chose to write.
The result separates diagnostic usefulness from evidentiary value. A trace may help identify a failure while remaining unable to prove what happened, particularly when the component under scrutiny is also the sole recorder.