Pre-registration: frontier language models fact-checking author-written abstract statements against the article's main text (AJHG, 2025-2026)

Pre-registration of a study on language models as fact-checkers of claims about research articles. This deposit freezes the protocol, the test set and the code before any evaluated model is called on the test set. The protocol is in Portuguese; README.md in the zip summarises it in English. Question. Given the main text of a research article (without its abstract), do four frontier language models (Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon, reasoning effort "low") (1) catch errors planted in statements written by the article's own authors, (2) avoid "correcting" statements that are right, and (3) answer "cannot verify" when the supporting evidence has been removed from the text? Test set. 70 research articles and brief reports from The American Journal of Human Genetics (CC BY 4.0, full text in Europe PMC, 2025-2026). Each abstract is split into sentences by code; 299 sentences with an anchor that code can alter and check (number, gene, HGVS variant, zygosity, inheritance, direction or certainty term) are scored, and each has one error planted by code and approved by the author. No language model wrote, altered or checked any statement, error or answer key. Two conditions: the full main text, with half of the scored sentences altered in each of two versions; and the same text with the passages that support half of the eligible sentences removed. Outcomes. Detection, false alarm and false correction under removed evidence (primary); abstention, error type, correction accuracy, real abstract-body divergences confirmed by blinded adjudication and the proportion of abstracts with at least one, consistency between two repetitions, refusals and cost (secondary). Files. PROTOCOLO-v2-frases-dos-autores.pdf (protocol); claim-check-ajhg-v2-preregistration.zip (protocol source, code, test set with the exact text of every call, article XML, the author's review of the planted errors, a functioning test outside the test set, and MANIFEST-sha256.txt with the SHA-256 of every file).