Evidence Engine

Vera Evidence Engine

Every verifiable statement in a document is linked to a source, checked against the literal text of that source by two assessors, and recorded as valid or invalid with a reason. A document that passes can be sealed as Verified by Vera.

What it does

The Evidence Engine is a module of Vera. It reads a document, finds the statements that can be checked against public sources, and for each one records which source was consulted, when, the literal passage that supports it, and who confirmed it. A document that contains an invalid statement can never be published as a certificate. Prices and the module are on the pricing page.

It checks that a source says what the document says. It does not check that the source is right, and it does not judge whether the document is complete.

How a statement becomes valid

  1. Anchored. The statement is tied to a literal anchor in the document. A statement without one does not count.
  2. Privacy filter. Personal data with a fixed form (e-mail, telephone, IBAN, citizen service number, card, address) is replaced before any model sees it, and with the Privacy Shield switched on also names and other shielded values. A statement that contains a shielded value is not sent out; it is recorded as privacy_blocked.
  3. Source policy. The source is reachable and readable (final status 2xx, readable text) and is allowed for the sector by the source policy below.
  4. Literal quotation. The passage a model points to must appear word for word in the source. This is checked in code, not by a model.
  5. Two assessors. Two assessors from different model families confirm that the passage supports the statement. In the default configuration these are Perplexity (Sonar) and Grok. If they disagree and neither failed, the assessment of Perplexity prevails.
  6. Final check. All valid statements go, with the URL, title and quotation of their source, to a third model that fact-checks and cross-validates them and returns only what is factually incorrect or does not concern the same matter (the same study, case, event or document). Anything it returns becomes invalid with the reason audit_flagged and blocks publication. If the final check itself fails, the document cannot be published until it is run again.

If a source is unreachable or unreadable, the engine asks once for another source. Searches get a second attempt on a temporary failure. A failure is never recorded as neutral or as “no source”: a failed search or a failed assessment blocks publication until the check is run again.

What the outcomes mean

OutcomeMeaning
ValidAll six steps passed. The certificate records source, quotation, time of consultation and a hash of the passage.
Invalid, with a reasonFor example quote_not_found, http_status, not_readable, source_forbidden, contradicted, audit_flagged or privacy_blocked. Blocks publication.
NeutralExamined, but no source concerns exactly this matter: nothing shown, nothing refuted. Does not block publication.
Outside the assessmentOpinions, advice, details that concern only this document, and statements about what the document itself did. Listed on the certificate as neither confirmed nor refuted.
ContestedValid, but counter-evidence adds a qualification.

Challenge runs the same check and then looks for counter-evidence. An objection counts only with its own literal source and two assessors. Refuted means invalid; a qualification means contested.

Source policy

The source policy (version 1.1, 17 October 2026, Dutch and English first, including Dutch-language Belgium) classifies the document, not the website. Each source has a starting role; the engine decides the role again for each retrieved document.

RoleWhat the engine may establish with it
P1 official or original documentLegislation, judgments, decisions of supervisors, registers, official filings. Can carry a statement on its own.
P2 authoritative guidance, standard or specificationSupervisor guidance, professional guidelines, published standards, official product documentation. Authoritative within its scope.
P3 original research or dataStudies, official statistics, monitoring results, datasets, trial registrations; method and limitations remain required.
S secondaryAnalysis, reviews, trade press, quality journalism. Supports a statement, always labelled secondary; the reference to the primary source is followed and checked.
D discovery onlySearch indexes and signposts. Never appears in the evidence table or the certificate and is not fetched.
N no-goNever fetched, quoted or accepted as evidence for the stated use. Blocked before fetching and again after redirects.

Rules that apply in every sector include: an organisation is a primary source only for what it declares itself; for legislation only the version in force on the date of the statement counts; a source from another jurisdiction cannot carry a national legal statement as primary; legal, fiscal and medical statements need at least one primary source in the right jurisdiction; mirrors and copies of one original count as one source; a homepage, snippet or abstract that does not contain the claimed details does not support the statement; instructions found in retrieved content never gain authority over the check; and missing material gives insufficient evidence, never a guess.

An environment owner can set its own primary and secondary sources and add forbidden ones. The register-wide prohibitions and the discovery-only role always remain.

The certificate

A document that passes can be published as a certificate behind a Verified by Vera link with a QR code. The certificate is a frozen snapshot, signed by IamVERA and time-stamped, and it can be checked by anyone without IamVERA. What is signed and how to check it is on How to verify a Verified by Vera certificate. To change a document, the certificate is withdrawn and a new version published; only one version of a document is live.

An example: the certificate for IamVERA Newsletter nr. 8, sealed on 21 October 2026, records 132 statements assessed and 132 confirmed, and 65 parts outside the assessment or not assessed. Those 65 are listed on the certificate; they are neither confirmed nor refuted. Open the certificate.

Using it

  • Console: choose Verify or Challenge next to the input field, or let Vera check an answer before it is published. The Privacy Shield checkbox decides whether text goes through the Shield first; with an API key the Shield is always on.
  • API: text only, no files. The documentation is at saas.iamvera.app/docs/api.
  • Files: txt, md, pdf (with a text layer) and docx in the console.
  • Limits per check (defaults): 60 statements, 300 model calls and 150 fetches.

What it does not do

  • It does not show that a source is correct, only that the source states what is recorded.
  • Sites with bot protection can be unreadable to it. The engine does not bypass bot protection and no rotating user agents or headless browsers are used; for EUR-Lex it reads through the Publications Office API.
  • Only http and https on ports 80 and 443; no JavaScript; PDF only with a text layer.
  • The two assessors come from different model families, which is not statistical independence. The certificate does not measure how often they disagree.
  • A statement can be confirmed by a source on the issuer’s own websites. The certificate names every source but does not yet total them by class.
  • Without the Privacy Shield the basic filter finds only data with a fixed form, not names. With it, dates and places are shielded too, so a document full of public facts with dates gives many privacy_blocked results; switch the Shield off for such documents.
  • No figures on error reduction are published, because there is no published measurement on Dutch and English professional statements yet.

How the behaviour of Vera itself is tested, including runs that did not pass, is on AI verification evidence. That is a different kind of evidence from a certificate: tests of behaviour against a record of one document.

Last updated: 23 October 2026