FinMirror Agent Lab · deterministic replay

Did the agent read what it cited?

Final-answer accuracy cannot establish a valid evidence path. FinMirror replays content-addressed document-read receipts, then checks retrieval claims, citations, operands, and the terminal calculation or abstention.

System	Answer accuracy	Verified trace	Correct but unverified	Gate
evidence-program-with-receipts	100.0%	100.0%	0	PASS
identical-output-without-receipts	100.0%	0.0%	126	BLOCKED

Interpretation boundary. This audit verifies that an observable trajectory is consistent with the exact evidence world. It does not claim access to hidden chain-of-thought and does not prove that the model lacked another information channel.

Self-contained artifact · no telemetry · embedded machine-readable evidence