biosingularity

Fact-check biomedical AI — with a receipt behind every verdict.

Paste an AI answer or a biomedical claim. biosingularity finds the checkable parts — cited papers, gene, drug and disease claims, DNA variants — and returns one verdict: pass, review, or reject. No model is in the verdict path, so it's deterministic and cannot hallucinate.

Built for teams shipping biomedical AI · researchers checking references · journals & research-integrity · pharma & biotech.

100% precision on the ACMG frequency rule behind reject — 5,179 common variants flagged, 0 truly-pathogenic wrongly rejected.
F1 0.84 · recall 0.97 in-silico consensus, used as supporting evidence only — never a standalone verdict.
0 models in the verdict path — deterministic; the same input always yields the same verdict.
See the benchmark →
Example — this answer cites a retracted paper and a black-box drug (expect REJECT). Edit and re-run with your own.

Demo depth: references · gene–disease · drug — full variant / interaction depth via CLI, API, MCP →

pass / review / reject — every finding links to the exact source record. Nothing is asserted grounded without a receipt.

Check anything — a DOI, gene, drug, trial id, or a claim

Or pick a specific check:

How it works

Three steps, no model in the verdict path — so the same input always yields the same verdict, and it can't invent a "looks fine".

1

Find the checkable parts

Pull out cited papers, gene / drug / disease claims, and DNA variants from the answer.

2

Check each against a primary source

Resolve every claim against ClinVar, gnomAD, Open Targets, ChEMBL, PubMed and more — deterministically.

3

Return one verdict — with receipts

Pass, review, or reject, and a receipt (a link to the exact source record) behind every finding.

The cardinal rule: a false green is the worst failure. Anything it cannot verify is held for review — never waved through. See the per-check rules →

What it checks

Every claim type is verified against authoritative primary sources — the identifier behind each finding is always shown, so a source key never reads as prose.

Claim typePrimary sourcesWhere it runs
References & retractions PubMed · 74,273-DOI retraction index (Crossref / Retraction Watch) demo
Gene–disease links Open Targets · DisGeNET · GWAS Catalog · STRING demo
Drug safety ChEMBL · FDA Orange Book · OpenFDA/FAERS · PharmGKB · DGIdb basic demo
FDA label facts new openFDA structured product labels — boxed warning · contraindications · dosage demo
Drug–drug interactions DDInter full depth
Variant pathogenicity ClinVar · gnomAD · dbNSFP · eQTL / GTEx full depth
Gene essentiality DepMap full depth
Gene–phenotype HPO full depth
Cell-line authentication now in demo Cellosaurus · ICLAC register (misidentified / contaminated) — 1,391 lines, bundled demo
Clinical trials ClinicalTrials.gov — stopped early? demo
Trial integrity new ClinicalTrials.gov — ICMJE prospective registration · FDAAA 801 results reporting demo

runs in the free browser demo partial in demo full depth (CLI / API / MCP over the 493 GB datalake)

The free browser check runs live PubMed plus a 74,273-DOI local retraction index (Crossref / Retraction Watch) — so it catches retractions before PubMed propagates them — and now the bundled Cellosaurus register, the FDA label sections, and the trial-registration arithmetic too. Full-depth checks (variant pathogenicity, drug interactions, adverse events, pharmacogenomics, eQTL) run over the 493 GB datalake via the CLI, the MCP connector, and the hosted API.

Who it's for

One verdict, three jobs — find yours.

AI / product engineers & agents

Verify before you assert — then repair

Gate every answer your model or agent produces before a user sees it, then get the corrected answer instead of a dead end: retracted citations deleted, contradicted sentences struck, sourced warnings appended. Repair only ever removes or annotates — it never rewrites a claim, so it cannot introduce a new one.

curl ".../api/verify?q=<answer>"
# → { recommend:"reject", claims:[…receipts…] }
curl ".../api/repair?q=<answer>"
# → { repaired_text, edits[], residual[], safe_to_ship }
Integrate: MCP + API reference →
Researchers

Check a manuscript or reference library

Scan a whole thesis or bibliography for retracted and contradicted work — free, no sign-up, with a clickable receipt on each flag. Then take the badge: a live 0 retracted shield for your README or preprint that re-checks on every view, so it turns red the day one of your references is retracted.

Journals & pharma

Screen at scale, with attestations

Verify a platform's generated discoveries or a manuscript batch and get signed, re-verifiable receipts for your audit trail. Now including the two integrity facts an editor acts on first: whether a reported trial was registered prospectively (ICMJE) and whether its results were ever reported (FDAAA 801) — plus misidentified cell lines.

Request a design-partner run →

Wire it into your agent

biosingularity is API- and MCP-first: your agent calls it before it asserts, then thresholds on one field — recommend. reject → block, review → don't auto-assert, pass → proceed. Every verdict carries a source receipt.

{ "mcpServers": { "biosingularity": { "command": "biosingularity-mcp" } } }
verify_answer repair_answer verify_discovery attest_answer + 15 more tools

Full MCP + API reference →  ·   Capabilities manifest (JSON)  ·   llms.txt

Get access

The demo above is free and needs no sign-up. For the full-depth hosted API, the MCP connector, the Python package, or a design-partner run on your own outputs, tell us who you are.

pip install biosingularity  ·  MCP connector available on request  ·  Free for academic use.