How we build the assurance layer.
Veridra functions as the evidence layer for other people's AI systems. We prioritize attestations that withstand adversarial scrutiny in legal and regulatory contexts.
Four operating principles guide every architectural decision: defensibility, auditability, human accountability, appropriate automation.
Four principles
Defensibility
Signed attestations must survive courtroom and regulatory challenges years after issuance. Implementation includes non-repudiable signatures, witnessed transparency logs, append-only schemas, and customer-held key management. A record signed today must be verifiable in a decade.
Auditability
Every claim the platform makes about a decision has to be independently verifiable without our cooperation. Verification tools are open-source, canonicalization follows RFC 8785, and customers can exit while verifying historical attestations using tools that don't require a Veridra account.
Human accountability
Material-impact decisions require named human owners rather than abstract system ownership. Article 14 oversight routes to identified reviewers, and approval actions are themselves signed evidence. "The AI did it" is not a defensible answer to a regulator; our platform ensures it's never the operator's answer either.
Appropriate automation
Safe processes are automated; consequential decisions demand human judgment. Risk determinations remain customer-controlled. The platform enforces technical mechanisms (cryptography, logging) while remaining deliberately neutral on customer risk preferences.
What we do not do
- Customer attestation data is excluded from model training.
- Explainability is sourced from customers, not internally generated scores that would falsely imply insight into their models.
- Corrections are logged as new signed events; records are never silently modified.
- Regulatory determinations remain customer responsibility — we produce evidence, we don't adjudicate compliance.
- Features launch only when failure modes can be explained to regulators, not when they merely pass internal QA.
Internal AI use at Veridra
Internal AI tooling (code assistance, documentation, support triage) is registered in the company's own Govern instance and subject to identical oversight standards as customer-facing systems. AI accessing customer data is customer-dedicated or disabled by default — we don't train our own models on the evidence we sign for you.