The input-cost angle seems especially interesting for a decision model, because repeated tool observations can dominate even when the model emits little or no text. Before changing the scorer, I’d try keeping the same search policy and applying a deterministic local compression pass to the serialized agent state, then measure decision agreement, search quality, and input tokens per decision. Contextpress (GitHub - Taha-azizi/contextpress: Deterministic context compression for LLM chat, RAG, and agent pipelines · GitHub · contextpress · PyPI) is a small way to test that boundary without claiming that compression is universally safe. If the decision model’s probabilities remain calibrated while duplicate observations fall, that would be a useful complement to the 162-session speed result.