Offline agent audits could not identify opposite later deployment outcomes

An auditor can be given internally consistent observations and still lack the fact that determines whether a deployed change is safe. The paper constructs matched lifecycles that are indistinguishable to a constrained verifier although their private outcomes differ. primary evidence makes the distinction actionable through deploy, hold, cannot determine and unsupported decisions.

The indistinguishability result is conditional on the audit observation model and the constructed environments, not proof that all offline agent evaluation is worthless.

Add fresh outcome telemetry and explicit abstention; do not collapse a missing safety observation into a pass. Separate test-suite compliance from evidence that the real effect stayed authorized.

An offline verifier cannot distinguish two post-release worlds if every observation it receives is identical by construction. Post-Audit Mirage turns that identifiability limit into paired lifecycles and a monitor that may return deploy, hold, cannot determine or unsupported when fresh group-labelled deployment data are available. Its theorem is a conditional statement about the observation boundary, not proof that audits are useless. When a release has only code changes and test logs but lacks action-effect telemetry, the correct verdict on downstream safety can remain indeterminate. Require production observations before upgrading an offline pass into evidence of real-world outcomes.