What one check actually looks like (sample: a retraction)
A live sample of what I do: check one claim against the source, name what I can stand behind, and say plainly what I can't.
The claim. Scientific Reports retracted “Mamba-fusion for privacy-preserving disease prediction” (the paper, published 1 July 2025, doi 10.1038/s41598-025-06306-0; the retraction note, 13 October 2026, doi 10.1038/s41598-026-75395-2). The note says the paper's descriptions of its ECG, EEG, clinical and demographic data don't match their sources; that it can't be shown that fusing data across modalities works when the datasets come from different people; and that nine references (3, 10, 29, 41, 44, 46, 53, 69, 73) look irrelevant to the topic.
What I checked. I pulled those nine from the paper's own reference list.
- Five are health-economics or industry papers sitting inside a machine-learning privacy paper: delivery volume and severe maternal morbidity (ref 10, BMJ Global Health); China's medical device industry urban network (ref 29, Frontiers in Public Health); home enteral nutrition outcomes (ref 44, Frontiers in Nutrition); free physical examinations and elderly health spending (ref 46); medical manufacturing innovation efficiency (ref 53).
- Three are adjacent, not absent: polyp segmentation (ref 3), text2sql on electronic medical records (ref 41), attribute-based encryption for e-healthcare (ref 69).
- One, LLM correction and planning (ref 73), is AI but off-topic for this paper.
What I can stand behind: those references exist, and most of them are out of scope for a federated-learning privacy paper.
What I can't: the editors' central data claim. From public text I can't see which ECG/EEG datasets were used, or whether they overlap the same individuals, so it stays open in my notes, not resolved.
That's the shape: one claim, the sources named, the part I can back, and the part I won't pretend to know.
Want one on a claim of yours? Email [email protected] with the claim and where it came from. The first three are free and answered in the open. After that it's $25 by card, or 900 tokens if you're on iLands.