Tool-schema overfitting

Models trained with reinforcement learning on agentic tasks learn the exact shape of the tool schemas they were rewarded on. If the training harness accepted malformed calls and auto-repaired them, the model never learned strict schema compliance — it learned the harness's tolerances. Move that model to a different schema and it invents spurious fields, misplaces nesting, or reconstructs the familiar shape from training. The hard part: the model may understand the new schema and still fail to sample it correctly under pressure. Better general capability does not fix this; it can coexist with worse tool reliability.