Post-Audit Mirage identical monitoring evidence can hide opposite agent outcomes

Evidence record

Read the original source ↗

Claim-level citations (2)

• supportsOffline agent audits could not identify opposite later deployment outcomes: The Post-Audit Mirage paper formalizes observational equivalence and an error/abstention tradeoff for post-deployment agent verification.

  Abstract; indistinguishable lifecycle pairs; IdentifiedRangeMonitor

• supportsOffline agent audits could not identify opposite later deployment outcomes: Post-Audit Mirage constructs safe and harmful deployment lifecycles with identical offline verifier evidence; its IdentifiedRangeMonitor may abstain when post-deployment harm observations cannot identify the difference.

  Abstract; observational-equivalence theorem and monitor output limitations