The distinction
Six questions that the word “verified” often collapses.
- 01Signed bytes. Which exact payload did the signature cover?
- 02Key control. Did the signature verify under the stated public key?
- 03Identity. Who or what is that key supposed to represent?
- 04Independence. Was the evaluator separate from the producer and free from conflicting incentives?
- 05Authority. Was that evaluator entitled to decide this question for this consumer?
- 06Truth. Is the underlying claim actually correct in the world?
Cryptographic verification directly addresses the first two. An external trust system may help with the third. The remaining layers require additional evidence and judgment. None should be silently inherited from a green signature check.
Why AI evaluations make this harder
A favorable result can rest on very different foundations.
An AI evaluation might record a formal predicate, a deterministic
heuristic, a model judgment or a self-assertion. Those are not
interchangeable merely because every native result says
supported: true.
The signature can preserve what the dossier declared. It cannot upgrade a model judgment into formal proof.
EvalDossier therefore checks both the cryptographic envelope and the
declared evaluation semantics. A formal predicate can produce a bounded
established status when its profile supports it. A qualitative model
judgment remains UNDETERMINED and
INCONCLUSIVE without stronger evidence.
Offline verification
Repeat the check without asking the evaluator again.
A portable dossier can carry the manifests, signatures, public keys and evaluation records needed for a local check. EvalDossier’s runtime makes no outbound request and requires no HREVN account or API key.
Offline repeatability applies to verification—not to regenerating byte-identical dossiers, recreating the evaluator’s internal process or proving that its observations were complete.
Where EvalDossier fits
It verifies a supported dossier; it does not evaluate arbitrary output.
EvalDossier is not:
- an AI evaluation framework or model benchmark;
- a general AI output validation service;
- a trust score, identity service or certification;
- a payment, escrow or settlement oracle.
It is a local verifier for supported signed evaluation dossiers. Its practical value is keeping the declared basis visible at the moment a consumer—human or agent—is tempted to rely on the result.
FAQ
Cryptographic verification questions
What does cryptographic verification prove?
It can establish that a signature verifies under a public key over specific bytes and that cryptographic commitments are internally consistent. It does not by itself prove that the signed claim is true.
Can a digitally signed AI evaluation still be wrong?
Yes. A signature can be valid even if the evaluator was mistaken, dishonest, incomplete or operating under a weak method. Signature validity and claim truth are separate properties.
Does a public key identify the evaluator?
Not by itself. A public key identifies a key. Connecting it to a person or institution requires an external identity and trust process.
Is EvalDossier a general AI output validation tool?
No. EvalDossier verifies supported signed evaluation dossiers. It does not validate arbitrary AI output, run evaluations or determine whether a model answer is correct.
Can cryptographic verification authorize payment?
No. A cryptographic result does not supply economic authority. EvalDossier fixes economicAction to OUT_OF_SCOPE.
Continue