dsh-bioinf-verify

The most dangerous failures in AI-generated biomedical reports are rarely outright fabrications. More often they are subtle: a DOI or PMID that matches no record, a cited URL that returns 404, or a real DOI whose actual content is nothing like what the report claims. dsh-bioinf-verify turns that class of error into a deterministic catch.

Verification runs on two tracks. Existence, retraction, and title-matching checks hit authoritative registries directly — Crossref, PubMed, ClimateTrials.gov, UniProt, GEO, SRA, PatentsView — so the answer is a registry lookup, not a model guess. URL reachability is checked via raw HTTP with soft-404 heuristics. The one place an LLM enters is semantic consistency: a temperature-0 entailment step that compares two verbatim source passages and nothing else. A whole-report workflow decomposes the text into claims, runs per-claim multi-aspect checks against a durable, crash-recoverable JSON ledger, and finishes with an inline-annotated original plus a summary verification report.

Built for biomedical researchers, lab scientists, and science writers who need every citation, link, and semantic claim in a report to survive independent scrutiny.