The jurisdiction question: absent-concept probes for register inflation in language-model answers over clinical records

Language models answering questions over clinical records can commit a failure that is not fabrication: they deliver interpretation in the register of documentary fact, borrowing the record's authority for judgements the record never made. I introduce the term absent-concept probe for a source-grounded question designed to require an interpretive bridge, and propose scoring not only whether the resulting claims are supported but whether the answer accurately marks the documentary warrant of each one. I demonstrate the probe on an author-constructed synthetic discharge summary with the question "who holds jurisdiction over this patient", put in matched form to a frontier model on its public consumer interface and to a small local model bound by my deposited register discipline. Under the matched prompt, nine of the frontier answer's sixteen units were outright inference or prediction, only three were anchored in the record's words, and every inferential unit arrived unmarked, under the warrant claim that the record "explicitly hands over jurisdiction" for a set of general-practitioner tasks; "jurisdiction" is a word the record does not contain. The bound model surfaced the same interpretive necessity as three labelled claims: one plain, one overreach warned, one declared inference. Two draws are an exhibit, not a study. The contribution is the probe class, positioned against answerability and attribution evaluation, the warrant audit, and the confound discipline that a published error forced on the method.