Your agent can find the policy, understand it, and still violate it. That happen

Finding the right policy did not prevent an agent from approving an invalid expense in HANDBOOK.md, a benchmark of simulated workplace tasks. In the paper's example, Anthropic's Opus 4.8 identified a $7,500 item requiring manager approval and checked who had approved it. It then incorrectly reinterpreted the junior analyst's identity as the financial controller and cleared the item. The failure occurred after the relevant evidence had been found. The paper documents the sequence.

The benchmark grades the final state of the simulated workspace with deterministic checks, including checks that prohibited actions did not happen. That makes the example useful for evaluating workplace automation: a correct explanation of a rule is weaker evidence than a record showing that the resulting actions followed it.