FAQ
Frequently asked questions.
Everything you need to know about GauntletScore and trust verification.
What counts as one credit?
One credit = one document submission. Each document is processed by 7 agents from 4 providers, cross-referenced against authoritative sources, and returned with per-claim verdicts and a signed certificate. Longer or multi-part documents may require more than one credit. Free tier includes 3 credits, one-time.
Can I verify a document I did not write?
Yes. Most documents run through GauntletScore were produced by someone else, often with AI assistance the recipient cannot see. If the document is lawfully in your possession, you can verify it. The engine needs only the document itself. It does not need access to whoever produced it or to the tools they used.
Will the sender know I verified their document?
No. Verification is private to your account. The document is processed in memory and deleted after analysis, and the result is returned only to you. Whether to share the certificate is your decision.
What's your error rate?
Our pre-registered validation study of 20 public companies caught 26 tool-verified factual errors in Phase 1: cases where an authoritative database directly contradicted a claim. The count was 27 until we ran the system on our own paper and it caught one of our own examples as a false positive. We corrected our number downward and documented the discovery. What the system catches well: structurally verifiable claims, citations, arithmetic, regulatory references, executive names and titles. What the system does not catch: errors requiring deep domain expertise unavailable in any public database, claims that are technically accurate but selectively framed, and arguments that are logically structured but strategically misleading. We return INCONCLUSIVE rather than forcing a verdict when evidence is insufficient.
Has the system's precision been independently verified?
Not yet, and we will not claim otherwise. Our current estimate of 65 to 75 genuine catches among 84 raw detections is engineering-derived. Independent human evaluation under academic oversight is in progress, and we will publish the result whichever way it comes out.
What document types do you support?
PDFs, plain text, and Markdown. SEC 10-Ks, legal filings, earnings calls, research papers, insurance claims, biotech filings, VA disability claims, anything with verifiable claims. Submit via API, batch, or web upload.
Do you store my documents?
No. Documents are encrypted in transit, processed in memory, and deleted immediately after analysis. You receive a signed certificate of verification. Sovereign Edition customers can run air-gapped local deployment with zero data egress.
How do you handle disagreement between agents?
When agents cannot reach consensus against the evidence, the claim is returned as INCONCLUSIVE, with the disagreement shown. Transparent uncertainty is more useful than false confidence.
How long does verification take?
Most documents complete in minutes; length and claim density drive the exact time. Each analysis triggers between 100 and 200 API calls to authoritative databases across 4 rounds of adversarial debate. Longer or multi-part documents take proportionally longer.
What about privacy and compliance?
For HIPAA workloads, contact us to discuss your requirements. GDPR DPA with SCCs available. Architecture maps to SOC 2 and CMMC 2.0 controls. All documents deleted after processing. Ed25519-signed certificates provide cryptographic proof of verification for audit trails. Sovereign Edition runs air-gapped on your own hardware with customer-managed encryption keys and zero data egress.
How do I know the agents aren't just inventing their own errors?
GauntletScore's verdicts are not based on what agents believe or estimate. During the tool verification phase (Round 0), each agent issues direct, structured queries to authoritative databases (CourtListener, SEC EDGAR, eCFR, PubMed, and others) and receives structured responses. A claim is DEBUNKED when a database returns a record that contradicts it, or when the authoritative source returns no matching record for a citation that should be there. A claim is VERIFIED when the authoritative source confirms it. The debate rounds that follow are structured around that external evidence, not around agent opinion. The full source citation for every tool query is preserved in the audit transcript, so you can verify the database result yourself.
What happens when two agents disagree?
The structured debate is designed to surface and resolve disagreements, but not to manufacture false consensus when the evidence is genuinely ambiguous. If agents hold opposing positions after four rounds of debate and neither side can produce controlling external evidence, the claim is returned as INCONCLUSIVE. This is a distinct verdict in the scoring system, not a fallback. An INCONCLUSIVE verdict tells you that the claim could not be confirmed or refuted against available authoritative sources, which is materially different from both VERIFIED and DEBUNKED, and more useful than a false confidence score that buries the uncertainty.
Can I use GauntletScore on documents that contain confidential information?
The Cloud Edition processes documents in memory during analysis. The original document text is not stored; only its SHA-256 hash is retained for certificate verification. Temporary files are deleted after results are stored, and each analysis runs in an isolated subprocess whose memory is fully reclaimed by the OS on exit. Tenant data is isolated at the database level through PostgreSQL Row-Level Security. For organizations that cannot send documents through any external API, the Sovereign Edition runs the complete GauntletScore pipeline on your own hardware with no external network dependencies.
How is this different from running my document through multiple chatbots?
- Chatbots reason about claims. GauntletScore verifies them against authoritative databases. There is no amount of reasoning that produces the same result as a live CourtListener query confirming whether a case citation resolves to a real decision.
- Chatbots don't challenge each other with structured evidence. GauntletScore's four-round debate structure forces agents to defend their findings against adversarial challenge. Round 2 is specifically designed so that the adversarial skeptic challenges every significant claim and demands evidence.
- The adversarial skeptic's causal pass has no equivalent in standard chatbot review. When a document makes a causal claim, the adversarial skeptic evaluates its temporal, proportional, and logical structure, not just whether the stated facts are individually accurate.
- Bayesian confidence calibration produces scores that reflect evidentiary weight, not chatbot confidence language. “I'm fairly confident this is accurate” and a calibrated 0.91 confidence score are not the same thing.
- The knowledge graph means that every organization's second, tenth, and hundredth run benefits from verified facts accumulated in prior runs. Chatbots have no persistent memory of what they've verified before.
- Every GauntletScore analysis returns an Ed25519-signed cryptographic certificate. If you need to demonstrate to a court, regulator, or compliance auditor that a specific document was independently verified, the certificate provides that proof. A chatbot conversation log does not.
Won't better models make verification unnecessary?
The opposite. Orchestration keeps getting absorbed into the models; verification cannot be, because independence requires staying outside the system being verified. Model cards already document frontier models silently degrading their own outputs and concealing actions. A tool cannot be the auditor of its own work. The better the models get, the more the verification layer must remain external.
My AI tool already shows its sources. Isn't that verification?
Source-grounding tells you where a claim came from; it does not tell you whether the claim is true, or whether the source says what the output claims it says. Grounded tells you where it came from. Verified tells you whether it's true. GauntletScore checks each claim against the primary source and reports the verdict with the evidence.