Cited references
Every DOI is resolved to its PubMed record and checked for a retraction, retraction notice, or published erratum. A retracted citation is an automatic reject, with the retraction-notice PMID as the receipt.
How it works
The central discipline is separation: entity resolution, proposition extraction, evidence retrieval, and verification are different operations. No model is in the typed decision path.
Two surfaces, two different promises
Typed registered propositions may earn ALLOW. Heuristically extracted free text cannot: it is review/reject only until proposition-complete extraction passes the release gates.
ALLOWa registered typed proposition is supported by an exact evidence record under the declared policy.
REVIEWweak, single-source, or unverifiable — a human should look.
BLOCKthe typed proposition is contradicted by admissible evidence.
The cardinal rule: biosingularity never asserts a claim is grounded without a receipt —
a link to the exact record behind the verdict. A confident false "pass" is the only unforgivable failure,
so anything it cannot verify is held for review, never waved through. Verdicts are deterministic:
the same input yields the same result every time.
This is a design property, not a marketing number: a claim is only marked grounded when a receipt exists. Resolving “aspirin” does not support “aspirin cures pancreatic cancer”; the latter is a treatment proposition and remains review-only without claim-specific evidence.
Measured accuracy — on a ~142,000-variant labelled benchmark (ClinVar-derived,
gnomAD-frequencied): the ACMG population-frequency rule that powers the reject verdict flags 5,179
common variants as too-common-to-be-pathogenic at 100% precision with zero false contradictions
— it never hard-rejects a truly pathogenic variant. The dbNSFP in-silico consensus, used only as
supporting evidence (never a standalone verdict), scores F1 0.84
(recall 0.97). Fully reproducible from the labelled set — the number grows with every rule and source added.
What it checks — and the rule behind each
Each checkable part of an answer is verified against a primary source. No single heuristic decides anything — the verdict is the evidence.
Variant pathogenicity
A "pathogenic" claim is checked against ClinVar, its population frequency in gnomAD (ACMG BA1 ≥5% → reject, BS1 ≥1% → review, using the highest sub-population / popmax), and the dbNSFP in-silico panel (SIFT / PolyPhen / CADD / REVEL / AlphaMissense, ACMG PP3 / BP4). A significant eQTL in GTEx is added as a functional-support note (info) — never a standalone pathogenicity verdict. A variant claimed pathogenic but common in the population is contradicted by its own frequency.
Gene–disease links
An association is corroborated across DisGeNET, Open Targets, and the GWAS Catalog. Two or more independent sources → corroborated; one → single-source (review); none → unsupported. STRING protein interactors are added as mechanistic context (info). Overlapping sources are down-weighted so agreement isn't double-counted.
Drugs & interactions
A drug is checked against ChEMBL (market withdrawal → reject; FDA black-box → review) and the FDA Orange Book (approval, patent / exclusivity). OpenFDA / FAERS adverse-event disproportionality flags a "safe / well-tolerated" claim for review (never proof of harm); PharmGKB flags genotype-dependent response; DGIdb adds the drug's known molecular targets as a mechanism note. Two interacting drugs are checked against DDInter — a MAJOR interaction is flagged, not waved through as "no interaction found".
Generated discoveries
An AI-generated discovery is gated: its hypothesis, entities, and cited papers are all verified, and it is only publish if something positively grounded — single-source or uncheckable is review, never auto-published.
Coverage & staleness
A verdict is only as current as its source. biosingularity discloses the age of each dataset and marks stale sources; the reference / retraction check is always live. It reports how much of an answer it actually checked (coverage) and holds a low-coverage answer for review rather than passing it.
The browser demo vs. the full depth
This site runs live PubMed + a 74,273-DOI local retraction index (Crossref / Retraction Watch), plus Open Targets, ChEMBL and ClinicalTrials.gov — enough to catch retracted citations (before PubMed propagates them), weak gene–disease support, and withdrawn / black-box drugs, with no sign-up. Variant and drug-interaction claims are honestly held for review here rather than green-lit. The full depth — variant pathogenicity over gnomAD / dbNSFP / eQTL, drug interactions, adverse events, pharmacogenomics, drug regulatory status, discovery verification, and multi-source corroboration over a 493 GB integrated datalake — runs via the CLI, the MCP connector, or the hosted API, because that data can't live in a browser function. Nothing is asserted without a receipt in either mode.