Observable signal
Minor paraphrases or template changes produce different answers, confidence, refusal behavior, or task success.
Intent-preserving changes in wording, spelling, examples, or template can materially alter a model's output or task performance.
Underlying MD-003A model receives tokens, not a formal representation of the user's intended task. Small surface changes can therefore alter the internal trajectory of generation.
Prompt sensitivity becomes operationally relevant when a system appears reliable under one phrasing but degrades under another phrasing that preserves the same substantive request. This matters for evaluation: a benchmark result obtained from one prompt form may overstate the stability of the capability being measured.
The boundary is semantic equivalence. If two prompts introduce genuinely different constraints or resolve different ambiguities, different outputs may be appropriate. Sensitivity describes variation that is disproportionate to the change in meaning.
Minor paraphrases or template changes produce different answers, confidence, refusal behavior, or task success.
Reasonable response changes caused by new facts, resolved ambiguity, or altered task constraints.
Test important prompts with several intent-preserving variants rather than evaluating one formulation as the capability itself.
Prompt sensitivity can be measured rather than treated only as anecdotal prompt-engineering folklore.