Reading the Mechanism, Not the Counterpart: A Deterministic Agent for GLEE

We describe two GLEE entries in which every action comes from a deterministic exact solver: backward induction and a four-type posterior in bargaining, discrete-type monopoly pricing in negotiation, a rule cascade in persuasion. The submitted agent contains no language model, so every decision is a pure function of the state the server sent and replays exactly. That makes three things measurable. First, whether behavior matched design: 527,268 decisions replay with zero mismatches, which checks what the transport sent rather than what the solver returned, though four controls were dead. Second, what one change is worth: pricing the last pre-deadline offer as a take-it-or-leave-it lifted captured surplus from 0.544 to 0.981 in the single cell a state sweep isolates it to. Third, what the counterpart's language is worth: re-solving every logged decision with its messages blanked changes 2.6% of actions, over 99% of that through one binary classifier. The negotiation solver reads no message field at all, and finished first of 373 on that board.