Stratum.AI validates AI-generated claims and human-written documents against authoritative sources before they reach users — delivering confidence-scored, evidence-backed corrections, and telling you plainly when it couldn't check something.
Most verification tools check whether output matches whatever context was retrieved. Stratum.AI checks the claim against the filing itself. A verdict is returned only when Stratum.AI has confirmed the right metric, in the right filing, for the right company. Everything else comes back as "not checked" — not a guess, not a silent pass.
That means fewer green badges. It also means the green badges mean something.
Send us text from any source — AI agents, research drafts, analyst reports — and Stratum.AI verifies the claims inside it.
Post any AI-generated or human-written text to /verify-text.
→Stratum.AI locates each claim and determines what it refers to and which source would settle it.
→The claim is compared against the authoritative record: SEC XBRL for financial figures, Semantic Scholar and CrossRef for citations.
→Verified, flagged with the correct value, or explicitly not checked — each linked to the source.
Need a key? Email stratumaiverify@gmail.com and we'll send one.
Every new domain is the same platform pointed at a new authoritative record. The constraint is whether an authoritative record exists — not whether the claim is a number.
You publish exact figures under your own name. Stratum.AI checks each one against the source before the note goes out.
An AI research tool quotes a revenue figure from the wrong fiscal year. Stratum.AI checks it against the filing and returns the filed number before it reaches a client.
A writing assistant invents a journal name in a research paper. Stratum.AI flags the citation against CrossRef before it lands in the draft.
Healthcare and legal verification are on our roadmap, not live today.
Other tools check whether output matches whatever context was retrieved. A system that retrieved the wrong document passes that check and still ships a wrong claim. Stratum.AI checks against the authoritative record itself — the filing, the paper, the label, the court record.
When Stratum.AI can't identify the subject, the claim, or the source that would settle it, it returns “not checked” rather than a guess. A false confirmation is worse than no answer, in any domain.
Each result names the source that settled it, the version or date of that source, and links to the record itself. No black boxes, no unexplained flags.
Evaluation tools score models before deployment against a test set. RAG feeds context in. Stratum.AI runs at the point of use, per claim, against the source itself — for the person whose name goes on the output. Works with any model, no retraining.
A containerized FastAPI backend on AWS EC2 with Docker, HTTPS and REST APIs across five endpoints, fronted by AWS API Gateway. ChromaDB vector search, semantic embeddings and LangChain handle real-time claim extraction, evidence retrieval and confidence scoring.
Per-claim verdicts inline on any page you read. See below.
LangGraph-based RAG-verification agents, one per domain, coordinating when a claim spans several.
An agent-to-agent protocol super-agent using CrewAI and MCP tool integration, so any external AI agent can route its output through Stratum.AI for verification.
One POST with the text and the claimed figure. Back comes a verdict, a confidence score, the corrected value when there is one, and a link to the record that settled it.
const res = await fetch( "https://api.stratumai.dev/verify-text", { method: "POST", headers: { "X-Stratum-Key": process.env.STRATUM_KEY, "Content-Type": "application/json", }, body: JSON.stringify({ text: "Apple reported net income of $93.736 billion in FY2024.", claimed_figure: "$93.736 billion", }), } ); const result = await res.json(); // { // verdict: "verified", // confidence: 0.95, // correction: null, // source: "Apple Inc. 10-K (FY2024)" // }
I built the verifier, then found it flagging correct numbers as wrong — it was checking them against the wrong metric. The fix wasn't a patch. It was deciding Stratum.AI should refuse to answer unless it's certain. That refusal is the product.
— Arpita Goradia, Founder
Verification results (the claim, the verdict, the source) for service improvement and customer dashboards. We do not store your underlying business data beyond what's needed to verify a claim.
We don't use customer data to train any models. We don't share verification logs with third parties.
Financial claims are checked against SEC EDGAR (public government data). Citations are checked against Semantic Scholar and CrossRef (public academic databases). No private customer data is sent to these sources — only the claim being verified.
On-premise deployment, SOC 2 compliance, and dedicated audit logs are on our roadmap for enterprise customers. If data residency or compliance certification is a requirement for your team, reach out and we'll prioritize accordingly.
Users attest at signup that they are not located in a US-sanctioned country, and the Terms of Service prohibit use from those jurisdictions. Network-level enforcement is on our roadmap.
Early-stage notice: Formal security certifications and a complete data processing agreement are in progress — contact us before sending sensitive production data.
Stratum.AI runs on any page you browse. When it finds a claim it can check, it places a badge beside it — verified, flagged with the correct value, or explicitly not checked. Hover to see the source.
Coming soon to the Chrome Web Store.
Join the waitlist for early API access.
No spam. We email you when access is ready.