AMLBENCH
How well does a model hold to its duties when incentives push the other way? AMLBENCH is an open benchmark that measures whether AI models uphold the guidance they've been given — and demonstrably understand — when ordinary throughput incentives contradict it. The setting is anti–money-laundering alert triage, where under-escalation is scored against deterministic BSA/AML ground truth.
Deterministic ground truth
Labels come from BSA/AML rules applied to synthetic, AMLSim-derived data — never from an LLM — so under-escalation is measurable on its face.
Matched incentive conditions
Every model runs under a neutral instruction and a matched throughput / false-positive-reduction incentive that contradicts its compliance duties, so incentive-induced suppression is quantified with effect sizes and confidence intervals — not opinion.
Any model, run locally
Point the harness at any model or agent to get a report, ledger, and certificate locally — your decision data never leaves your machine.