Skip to content
Devernance

Methodology

How we map and interpret.

Evidence and readiness you can reproduce: a pinned corpus, mappings a person accepts, and checks that fail closed.

  • A pinned legal corpus

    Systems are read against a fixed, versioned copy of the legal text from EUR-Lex, including the Digital Omnibus on AI. The pin is the text the classification reads, so a result can be reproduced later.

  • Mappings are proposals

    Links between a system and the obligations that may apply arrive in a queue. A person accepts or rejects each one. Nothing is applied automatically.

  • Control modes

    Each control on a provision in force has a mode: INFORMATIONAL, WARNING, APPROVAL_REQUIRED or BLOCKING. The mode decides whether a gap warns, waits for a person, or stops the release.

  • Fail closed

    Missing evidence, a failed evaluation or an open data-contract breach never passes. If the decision cannot be read, the gate returns BLOCKED.

  • People decide

    Risk class is never changed automatically. Engineering, compliance and legal sign off, and every decision is written to a hash-chained ledger.

  • Checkable outside the platform

    Evidence exports can be scanned with the open-source evgraph-cli. The evidence pack and the scan report the same gaps.

Limits

What we do not infer.

  • We do not invent a missing approval timestamp or dataset licence. The finding stays inconclusive.
  • We do not decide whether a system meets the law. Counsel makes that determination.
  • We do not certify systems or act as a notified body.
  • We do not send your content to a language model by default.