Deterministic Audit of LLM-Generated Research Claims

Large language models are increasingly used to draft the narrative parts ofresearch: reports, verdicts, discussion sections. The failure mode that mattersis not obvious fabrication but \emph{plausible overstatement}: the number isright, the interpretation is not, and the prose is confident. We describe aprotocol that makes such overstatement mechanically detectable, and report whathappened when we applied it to our own corpus of 1{,}416 probes across tenexperimental series. The audit covers ten $p$-values and twenty-one fractions,and Section~\ref{sec:status} states its scope. The protocol rests on six requirements: pre-registration freeze, external memoryinstead of context recall, separation of generator and auditor, deterministicparity checking of every reported number, a three-tier taxonomy of claim errors,and multi-model adversarial review. We contribute an executable checklist oftwelve items and report its honest status on our own work: six pass, one isamended, one fails, two are partial, two require human judgement. The protocolis not fully applied, and the checklist says so. The Russian version of this preprint is the second file in this record:main_PACKAGE_v2_ru.pdf. / Русская версия — второй файл записи:main_PACKAGE_v2_ru.pdf.