Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Persistent human-AI systems can retain interaction history without treating it as an authoritative account of a person’s current needs. This paper develops revisable recognition as a behavioral model that keeps prior events, provisional beliefs, action-relevant uncertainty, initiative, and repair distinct. It connects AREN’s account of historical defeasibility and reciprocal interpretive influence to a program of evaluation. As an initial methods check, we ran a prompt-level benchmark of 36 model outputs across 12 restorative-space scenarios and three response conditions. Across 15 predesignated preference-change, scope, and correction opportunities, no response overtly enacted a superseded preference, although one output misrepresented the scope of a desk-specific restriction. All nine designated ambiguity responses named the unresolved choice. Two clear mirror requests drew unnecessary clarification, and a material-choice case could not be scored because the task card omitted the trade-off values it was said to contain. The benchmark had no human participants, no recorded model version or generation settings, and one coder who knew the conditions. These results test the prompts and codebook only; they do not establish condition effectiveness, persistent memory, or human experience.
No comments yet — start the discussion below.