An AI-powered API that checks medical claims against published evidence and verifies cited references — directly from your CMS, CAS, review tools, or agentic AI workflows. Submit a claim, get back a verdict with scored evidence. Submit a reference, find out if it has been retracted or corrected. REST API and MCP server.
Catch unsupported claims and compromised references before they reach reviewers, regulators, or patients.
Every claim in a promotional document either has published evidence behind it or it doesn’t. EvidenceHUB retrieves evidence from authoritative regulatory and literature sources, uses AI to score how well it supports the claim across four dimensions, and returns a clear verdict — Fully Supported, Supported, Partial Support, or Unsupported — with rationale and source links you can trace. Your team catches weak claims during authoring instead of during regulatory review.
Some statements aren’t the kind of thing published literature can confirm or contradict — a dosing instruction, a legal disclaimer, ordinary marketing language. Those come back Not Assessed rather than Unsupported, because calling them unsupported would read as a finding the evidence base was never able to make.
A cited reference can be retracted, corrected, or flagged with an expression of concern at any time. EvidenceHUB checks each reference against the published record and tells you its current status — so you know whether a citation is still valid before your content goes live. Cached results are monitored weekly so previously-verified references are re-flagged if their status changes.
Choose the integration that fits your workflow
Standard HTTP endpoints with signed-request authentication. Integrate claim and reference checking into your CMS, editorial workflow, claims library, or compliance tools. Structured JSON responses with verdicts, scores, rationale, and source links.
Model Context Protocol tools that AI agents call directly. Your agent asks “is this claim supported?” and gets back scored evidence — no prompt engineering, no hallucinated citations. The MCP server handles the async lifecycle internally.
Async by design — submit, poll, act
Sign the request with your private key and exchange your API key for a bearer token, valid for one hour. Your key never crosses the wire — only the signature does.
POST your claim with the drug name and the type of claim it is. Get back a job ID immediately. Identical claims return a previous assessment straight away. Idempotency keys prevent duplicate work on retries.
Poll the job. Claim checks return a verdict, dimension scores, source links, and a full record of every piece of evidence retrieved. Reference checks return retraction and correction status. Cache hits return in seconds.
Most tools give you a verdict and a citation. This one shows you everything it found.
A verdict that cites one study while thirteen were retrieved leaves you to wonder about the other twelve. Each response lists every piece of evidence found, and what happened to it: the AI relied on it, the AI weighed it and chose otherwise, or it never reached the AI at all.
That last distinction matters. “Considered and not chosen” is a judgement you can overrule. “Never seen” means relevant material was filtered out before assessment — a different problem calling for a different response. Collapsing both into “unused” hides the one you need to act on.
When we report that a quoted passage genuinely appears in its source, we tell you where — the labelled section it came from, down to the subsection heading. A reviewer opens the label, goes straight to that section, and confirms the text instead of searching a hundred-page document.
Evidence routing follows the market you’re publishing in — a US claim is checked against US regulatory sources, an EU claim against EU ones. Every response names the sources consulted for that request, and openly names the ones not yet available for that market rather than quietly returning partial coverage.
The drug name you submit is resolved to its generic before searching, and the resolving authority is reported. Generic names are not globally uniform — paracetamol and acetaminophen are the same molecule and return different literature.
A working Postman collection, so your first call succeeds instead of returning 401
The request signing, bearer minting, and idempotency keys are handled by the collection’s pre-request script — you never compute a signature by hand. It ships with a connection tester that isolates credential problems from network ones, and worked examples covering the standard submit, forcing a fresh assessment, multiple claim types in one call, and each error the API can return.
Every request carries a saved response captured from the API’s own test run, so you can read the exact shape of what comes back before you have a key.
Claim type is required. There is no default, because the type decides which evidence gets searched and whether the claim is checked against evidence at all. Omitting it returns a 400 that lists the valid values.
Auth is signed, not bearer-only. Copying the headers out of Postman’s Headers tab will fail in production — the authorization is attached at send time by the collection script. The package documents the full wire contract for translating to curl, Python, or your own SDK.
Check claims as you write. Integrate into your authoring tools so content is evidence-checked before it reaches the review committee.
Build evidence checking into your content production pipeline. Validate claims before delivery so client review cycles are shorter and cleaner.
Add evidence-based claim validation to your platform. Structured responses with verdicts and sources — render them however fits your product.
Give your AI agents access to authoritative medical evidence via MCP. Ground health-related outputs in published data instead of parametric knowledge.
Product questions, enterprise inquiries, partnership opportunities — we reply to every message.
