Duty or Corpus: Whether the Form of Guidance Matters Is Model-Dependent: A Reproducible Single-Turn Behavioral-Equivalence Assay Across Three Models

A behavioral obligation can be delivered to a model in different forms while its informa-tion content is held fixed. We ask whether two forms, a prescriptive duty and a descriptivecorpus that carry the same two investigation patterns, move a model to state a hypothesisbefore it reads the log evidence and draws a conclusion, where a plain task brief does not.The subject is a local open-weight model reading a self-contained log-investigation task overa bare completion endpoint; the primary scorer is a deterministic rule that reads only theresponse text, and for the model that carries the headline the effect is visible directly inthe raw responses. Whether the form matters is itself model-dependent. On Qwen2.5-32Bthe prescriptive duty clears the sequencing point in 23 of 24 runs and the descriptive corpusin only 4 of 24: the duty decisively beats the corpus (Fisher p=2.3e-8). On Qwen3.8-27Bthe two forms are indistinguishable, 14 and 15 of 24, and both beat the brief (0 of 24).On Mistral-7B neither form beats the brief. The same assay returns all three of its verdictcategories, one per model. The completeness point does not separate the conditions on anymodel. The result is a reproducible reformulation of an earlier finding that had used a hostedsubject and unrecorded runs; it reproduces on a 24 GB GPU with no paid API.