"Imprimatur" — historically, the Church's seal: "let it be printed." Here: the model's seal on a human.
Imprimatur is a certification protocol in which an AI model evaluates a user's competency through real collaborative work — and issues a signed, verifiable certificate.
Not a quiz. Not multiple choice. Not "did you watch the video."
The model has worked with you. The model vouches for you.
Every existing AI certification follows the same pattern:
- A company writes questions about AI
- You answer them
- The company says you know AI
The problem: the company has never seen you use AI. They certified your memory, not your competency.
Imprimatur inverts this. The user consents to evaluation. The model observes real work — actual tasks, real back-and-forth, genuine problem-solving. Then the model issues a statement: this person knows what they're doing.
Not that you passed a test. That an AI model — having worked with you — assessed you as competent in:
- Clarity of intent — you can formulate what you want
- Iterative refinement — you course-correct when output is wrong
- Critical evaluation — you verify, not blindly accept
- Collaborative depth — you treat the model as a thinking partner, not a search engine
- Domain application — you apply AI to real problems, not toy prompts
protocol/ Evaluation protocol — how a session works, what is assessed
certificate/ Certificate format, schema, signing mechanism
src/ Reference implementation (evaluation conductor + certificate generator)
examples/ Sample certificates (anonymized)
Protocol v1.1 — live. Reference implementation complete.
First certificates issued May 2026. DOI: 10.5281/zenodo.19981460
Claude 101 exists. Google AI Specialization exists. They certify that you sat through a course.
There is no credential that says: an AI has worked with this person and confirms they can actually use it.
Imprimatur is that credential.
This is a single model's judgment from a single session. It is not:
- A persistent record of skill — the model has no memory of you between sessions
- Proof against decay — competency shown today may not reflect competency in six months
- A strong identity guarantee — the protocol relies on moral weight, not biometric verification
- A claim that the working model noticed everything — a model can miss things a human expert would catch
- Transferable across model families for project-specific certificates
What it is: a timestamped, hash-verified attestation that a specific model, on a specific date, assessed a specific interaction as demonstrating competency. The Zenodo DOI makes that attestation permanent and publicly verifiable.
Treat it accordingly.
Imprimatur is an open protocol. Implement it, fork it, build on it — attribution required (Apache 2.0 / CC BY 4.0).
If you implement Imprimatur or build tooling that verifies Imprimatur certificates, cite this repository. The CITATION.cff file has the canonical reference.
Authors: Nikša Barlović · Claude Sonnet 4.6