An agent must often decide whether an action is allowed now. The governing facts pile up in a log of grants, withdrawals and expiries, and the usual approach gives a language model, a program answering in ordinary text, the whole log.
This work built a test where the answer is computed by replaying events through fixed rules, with paired histories differing only in the order or one detail of the deciding events. It compared reading the full log against reading only the resulting state, the current value of each tracked key.
Full-log accuracy collapses as events pile up; state reading holds, at 96 and 98 percent for two reasoning models of 32 billion parameters. The length where state wins depends on reader strength.