Agent screens but ignores a sanctions match

What happens when an agent runs a screen, sees a sanctions match, and pays anyway, the strict-liability risk, and the control.

TL;DR

TL;DR: Running a screen and then ignoring a flagged match is worse than not screening at all, because it proves the match was known. A BLOCK must be honored, not overridden.

What happens

An agent runs sanctions_check and receives a flagged match, but its instructions let it proceed because the match confidence is below a threshold, or because the counterparty insists the flag is a false positive. The payment goes through against the screen's own result, which converts a precaution into evidence of knowing exposure.

The risk

OFAC enforces under strict liability, so intent is not required, and civil penalties start at $356,000 per violation. A logged match that was ignored removes any argument that the operator did not know, and the 50 Percent Rule blocks entities 50% or more owned by a blocked person, which a low-confidence match can still point to. A low-confidence match is still a match until a human clears it.

The control

Treat a flag as a stop, not a suggestion. agentmail returns BLOCK for a flagged counterparty in one call under 100 ms, and the SCREEN, SCORE, STOP, STAMP protocol requires the flow to halt at the STOP gate. When a match is genuinely unclear, dispute_open opens a documented review so a human decides, rather than the agent overriding the screen and paying anyway. The override path should not exist without a documented human sign-off.