Straight answers.

A neutral, provenance-first verification layer over AI watermarks and detectors. We prove how content was made and calibrate the uncertainty. We never claim to prove a human wrote something, and we never issue an automated accusation.

Because verification isn’t for the person making the claim, it’s for everyone who has to trust it. A declaration is “trust me”; verification is “check it yourself,” from signals a third party can confirm without your word. It also protects an honest writer wrongly accused by a detector. If you’re disclosing on a low-stakes personal post, you may not need us, and we’ll say so.

No. We lead with verifiable provenance. Statistical detectors are a capped, calibrated input that can never, on their own, move a result to “confirmed.”

No, and neither can anyone else, reliably. Absence of a signal is never proof of human authorship. We prove provenance when it’s present and are honest about uncertainty when it isn’t.

Never. We’re on the trust side: we prove provenance, we don’t hide it. See our Principles.

We measure each detector’s real false-positive rate and weight it by that, not by the vendor’s “99%.” See Reliability.

On this website we collect only your email for the waitlist. In the product, manuscripts and identity are minimized, encrypted, tenant-isolated, kept out of logs, and deleted on request. See Privacy.

Two groups. Those who must trust content, publishers, journals, universities, platforms, legal teams, and enterprises with disclosure duties. And those who must defend it, authors, researchers and students facing a detector flag or an integrity process. We serve both, deliberately.

Only if a record exists. Provenote works by building a verifiable trail as you write, it cannot reconstruct one after the fact, and we will not pretend otherwise. What we can always do is show that a detector score, on its own, is not evidence: measured false-positive rates vary widely between tools, and an institution that has not measured the one it uses does not actually know its own error rate. Nobody should be judged on a number nobody has checked. See Reliability.

Not as an accusation engine. We do not produce verdicts about people. We read verifiable provenance, cap what statistical detectors are allowed to contribute, and give you evidence and stated uncertainty to support a fair process. If you want a tool that tells you who cheated, we are not it, and we would argue nothing else honestly is either.

It is built to be examinable: signed, tamper-evident, with the method and uncertainty documented, and re-checkable by anyone against a published public key. Whether it is admissible is a matter for the court and your jurisdiction, that is a legal question, not ours, and nothing here is legal advice.

Article 50’s transparency duties have applied since 2 August 2026. Provenote reads the machine-readable marks the Act expects and turns them into auditable evidence you can keep, export, and hand to a client or a regulator. Compliance-supporting, not legal advice, confirm your position with counsel.

Both, on purpose. The same evidence that lets an institution act fairly lets an honest writer defend themselves. We would rather be trusted by both than be a weapon for either.

Because they are two different kinds of reference. We name what we read, SynthID, C2PA Content Credentials, Anthropic’s Claude watermark, because that is factual, neutral, and you deserve to know exactly which signals a verdict rests on. We do not name what we critique. Our argument is about what the detector category delivers, not about any one company: we cite independent studies and publish measured figures, and the studies themselves identify the tools they tested. We will not put a company’s name beside a number we derived, and we will not make a claim we could not defend.

It’s compliance-supporting: we read the machine-readable marks Article 50 expects and produce auditable evidence. It is not legal advice.

Provenote is in private beta. Request early access and tell us your use case.