A State-Machine Gate Compiled From τ²-bench's Airline Policy Lifts a 235B Agent From 0.39 to 0.54 pass^1, and an Accurate Checklist Loses to a Ledger the Agent Writes Itself

arXiv·high signalXBlueskyLinkedInCopy link

Zhang, Kweon and Han hold the model and rules fixed and vary how task state reaches the agent: raw transcript, exact checklist, per-turn directives from a compiled state machine, or an enforcement gate that refuses state-violating actions. Showing accurate state was unreliable, and an unverified ledger the agent wrote itself beat an accurate checklist it was shown. The gate raised a 235B agent's τ²-bench airline pass^1 from 0.39 to 0.54 and did nothing for a 35B agent that rarely broke policy. On cue-driven PM-Bench tasks the gate's wrong judgments pushed a 35B agent below its raw-transcript baseline. Enforce state in code only where failures are state-decidable and frequent.