VERSIO-AI is the audit tool that does. Paste a DOI; the pipeline reads the open-access full text when one resolves (otherwise the abstract), runs a single Claude Opus 4.7 extraction, and renders the result as one editorial card: model variant tested, evaluation date, elicitation completeness (reasoning, tools, prompting, the strongest within-family variant), and whether the paper's conclusion stayed scoped to those tested models or escaped into class-level claims about AI or LLMs.
Was the tested model named precisely enough that a domain reader can place it on the capability timeline?
Was reasoning enabled, were tools and search allowed, was prompting and context disclosed, and was the strongest variant of the chosen family generation actually used?
Did the conclusion sentence keep its subject scoped to the tested model, or did it generalise to AI / LLMs as a class?
How far behind the contemporaneous Arena-elicited frontier was the tested model when the evaluation actually ran?
Frontier Lag(the bibliometric audit this tool is the public face of) measures three structural disclosure failures in academic AI-capability research: tested models routinely months or years behind the elicitable frontier at the moment the evaluation runs (lag), comparator sets that span only a paper's preferred tier rather than the contemporaneous top (comparator inadequacy), and conclusion sentences that generalise findings from one model to AI / LLMs as a class without the cross-model breadth such claims require (frame asymmetry).
VERSIO-AI is what falls out when the same Opus 4.7 extraction prompt that scores the audit corpus (medicine, law, coding, education, scientific reasoning; preprint forthcoming) is re-pointed at a single DOI. Every verdict carries the verbatim substring that drove it; every frontier comparison anchors against the Arena trajectory and a registry of ~160 named models. Nothing on the card is a guess.