Post-training checkpoints for tool-use policy research. Each card states the measured behaviour, rejected checkpoints, and what was not measured.