Three manuscripts, one question

The two notes before this one set the problem up in general terms: what a summary costs, and what a reader does with a provenance mark. This is where the question went next, into three manuscripts.

Three papers are in review at the moment. Read separately they look like three topics: empathy in generated health text, pairwise judgement of model responses, and provenance labels. Read together they are one question asked at three points in the same pipeline.

First, what the model writes. If you prompt a system for a warm, patient-facing message, does the warmth come from the model or from the prompt? And when a reader calls the result empathetic, are they responding to the text or to the fact that they know who wrote it? The paper spends most of its effort on making those two separable.

Second, how we judge it. Most model comparisons reduce to a preference: A or B. That preference carries assumptions about identity and style that rarely get stated. The second paper treats the judgement itself as the object, including the part of it that participants cannot explain.

Third, how we read it. Even a correct provenance label lands in a reader with a deadline and a prior. The third paper asks whether the way a reader approaches the text decides more than the label does, which is the least technical of the three and the one I think matters most.

Production, evaluation, reception. The failure I care about happens in the seams between them.

Titles stay unpublished while review runs, and full texts go out on request. If a reader wants the detailed version of any of this, the contact page is the shortest route.