Atomic claims
Separates dates, quantities, attributions and factual propositions while preserving their source spans.
Mephistopheles breaks AI output into atomic claims, retrieves inspectable evidence, and reports support, contradiction, dispute, uncertainty, and coverage—without turning a score into a guarantee.
Separates dates, quantities, attributions and factual propositions while preserving their source spans.
Queries structured and public sources, then normalises canonical destinations and provenance.
Applies a versioned profile, including source-class, freshness, jurisdiction and abstention rules.
Shows what was examined, what may have been missed, and which material findings drive risk.
This demonstration calls the same versioned verification endpoint used by the workbench. Results are generated, not pre-written.
Retrieval and model providers may take a little time. Do not submit secrets, personal health information, or other sensitive content in this public demo.
Every verdict includes its limits and provenance.
It means the active policy obtained enough suitable evidence for the examined claim at the displayed time. It does not establish universal or permanent truth.
Contradiction, uncertainty, incomplete coverage and failure states remain visible. Material findings can be routed to a human review record without overwriting the machine result.
Map original passages to atomic claims, verdicts, evidence, corrections and review state.
Open workbenchValidate PDF, DOCX and text uploads, follow durable processing, and export an annotated review.
Review a documentVerify text you explicitly select through a stable Manifest V3 side panel with narrow permissions.
Review permissionsIntegrate the versioned result contract, asynchronous documents, webhooks and governed policies.
Read developer docsCoverage, risk and unresolved evidence are reported separately. A result remains qualified when material content was missed or suitable evidence was not obtained.
No historical benchmark percentage is presented as current until it reproduces from the recovered, versioned evaluation set under a frozen release configuration.
Known limitationsUploads are bounded, signature-checked and malware-scanned. Provider credentials and retained content use separate encrypted controls. Public documentation states what may leave the service boundary and what zero retention omits.
The product does not imply SOC 2, ISO 27001, HIPAA or other independent assurance without the relevant evidence and agreements.
Review subprocessorsUse typed Python and TypeScript clients, scoped keys, request IDs, rate-limit metadata, version discovery, owner-bound document jobs and replay-resistant webhooks.
Explore the APIThe anonymous demonstration is deliberately limited. Account and API usage expose metering, quotas and hard cost ceilings; final public prices remain gated on measured provider cost and margin.
Review plans and billing boundarySee exactly what was checked, what evidence was found, and where the system abstained.
Verify now