For universities, integrity offices, journals and publishers

Keep a record of how your work came together, as you write it.

If your authorship is ever questioned, you answer with evidence instead of your word. Each draft leaves a fingerprint, the fingerprints are chained so nothing can be inserted or reordered later, and the finished chain is sealed into a certificate anyone can re-check. Your words never leave your machine.

We are not an AI detector, and we do not read your work. We do not score how AI-sounding your writing is, and we never accuse anyone.

Looking for a watermark, passage by passagefound ▲
Each bar is one passage. A bar above the line carries a watermark we could read. A bar below the line means we found nothing, and that is not evidence a human wrote it. Most writing in the world carries no watermark at all.
How it works

Three steps, and no jargon.

STEP 1

You write as you normally do

In Word, Docs or whatever you already use. Each time the work is saved, a fingerprint of the text is taken on your own machine. The words are not sent anywhere.

STEP 2

The drafts are chained

Each fingerprint is linked to the one before it, so nothing can be inserted, removed or reordered afterwards. The record stays with the work, in your own storage.

STEP 3

It is sealed, and anyone can check it

One certificate at the end, re-checkable offline against a published key, years later. It shows what the record holds, including where there are gaps.

Where this stands. The chain and the certificate are built and working: a real six-month manuscript has been sealed this way and independently re-checked. What does not exist yet is a one-click way for a writer to do it alone. Today it runs inside an institution, or alongside us during the private beta. We would rather say that now than after a procurement process.

The longer version →   or   see a sample result →

What we check

Four kinds of signal. One of them fully works today.

The record above is one half of what we do. The other half is reading a file somebody has handed you, and here we are blunt: for ordinary written text it usually finds nothing, because no word processor writes a record and no conforming product declares a text media type. Most of this industry describes what it intends to do. We publish a table of what actually happens when you send us a document, including the checks that are not available to us yet. We would rather lose a sale than have you find that out later.

WORKING

Content Credentials

Signed records attached to a file. We read them and verify the signature.

PARTLY

SynthID

Google’s watermark for AI-generated content. We read it only where the keys are known to us, which is not Google’s production watermark.

NOT YET

Claude’s watermark

Anthropic’s watermark on text written by Claude. Detection is private preview and we have not been granted access.

CAPPED

Style detectors

Bring your own. Weighted by how often it is wrong, never an accusation.

The full table, and the three things we are not →

Bring your own detectors

Connect the detector you already pay for. We make it honest.

Plug in your own commercial detector API, alongside our native provenance signals, or on their own. Your key, your account, your data terms. We add the evidence a detector can't see, and we tell you how much to trust the score you're paying for, measured against our own held-out human text, not the vendor's advertised figure.

MODE 01

Native

Provenance-first: watermarks, C2PA, and our own calibrated signals. No third party required.

MODE 02

Bring your own

Your detector subscriptions, run through one report, each one calibrated to its real false-positive rate.

MODE 03 · DEFAULT

Hybrid

Native provenance and your detectors, combined, with disagreement surfaced, never averaged away.

Works with Major commercial detector APIsYour own endpointYour key, your account, your data terms

Every third-party detector stays a calibrated, capped input, it can never, on its own, produce an accusation.

Why now
2 Aug 2026
EU AI Act Art. 50 transparency duties, already in force
~750
Papers a university calculated would be wrongly flagged in a year, using the vendor’s own claimed 1% error rate
9%
Human text mislabelled as AI by OpenAI’s own classifier, which it retired for low accuracy
Neutral
Built to read provenance signals across providers, not to favour one, subject to each provider's detection access.

Every number here, with its source →

◈ Where we stand

Provenote is on the trust side. We verify provenance and calibrate uncertainty. We do not strip, weaken, or evade watermarks or detectors.

Read our principles →

◈ Checked by someone other than us

The C2PA Conformance Program lists Provenote as a conformant validator product, with a conformance date of 11 November 2026. The listing covers validating JPEG images, PNG images and PDF documents against C2PA specification 2.2.

It does not cover generating or signing Content Credentials, and it does not cover written text. We are not certified, approved or endorsed by C2PA, and we are not on any trust list: conformance means our validator was tested against the program’s requirements and met them. The record is public and machine-readable, so you can check it yourself rather than take our word for it.

First commercial customer

Arika Publications LLC, publisher of Brilliance for All, uses Provenote to keep a record of how its manuscripts are written. The first is sealed: 26 drafts between April and October 2026, re-checkable by anyone.

Arika was founded by Provenote’s founder, so we are not offering it as an arm’s-length reference. We name it because the record is real and you can check it yourself. See the sealed record →

Our entry on the Conforming Products List →

Who it’s for

What you get, depending on which side you’re on.

A record of how work came together, kept while it is written. Below is what that is worth to each side of the same decision: the institution that has to judge a submission, and the person who has to answer for one.

Universities & academic integrity

When you must make a decision that affects a student’s record.

Provenance-first evidence, plus each detector’s measured false-positive rate rather than its marketing number, so a case rests on something checkable and due process is documented.

For universities and schools →

Publishers & journals

When a submission arrives and you must decide what to trust.

An auditable provenance report per manuscript: content credentials and provider watermarks read for you, uncertainty stated plainly, and a record you keep.

For publishers and journals →

Students · school & university

When an essay you genuinely wrote gets flagged.

A verifiable writing record you build while you work, so a challenge is answered with evidence rather than argument. We never issue an accusation, and we tell institutions plainly that a detector score, on its own, does not establish that someone used AI.

Authors & researchers

When you submit work and your authorship may be questioned.

A tamper-evident record of your drafts, in order, with their dates, plus a signed certificate anyone can re-check offline. If a detector flags you, you answer with evidence instead of your word.

Corporate & enterprise

When you publish, procure, or carry disclosure duties.

Verification at scale through an API, plus exports you can use as part of your own evidence under EU AI Act Article 50, the transparency duties that require AI-generated content to be machine-readable and, in some cases, disclosed to a person, and support for your own detector keys. Compliance-supporting, not legal advice, confirm your position with counsel.

We serve the university and the student, the journal and the author. That is deliberate. A verification layer that only ever works for one side is not neutral, and neutrality is the entire product.

We build the record and read the signals, and we are honest about uncertainty when the evidence runs out. The one thing no tool can do, ours included, is look at a finished text and declare who wrote it. So we do something that works instead: keep the record of how the work came together, which is evidence an author can actually produce.

Provenote produces technical evidence and stated uncertainty. It does not render verdicts about people, and nothing here is legal advice.

Disclosure isn’t verification

“If I just say I used AI, why do I need this?”

Because verification isn’t for the person making the claim, it’s for everyone who has to trust it.

DECLARATION

“Trust me.”

A self-declaration can’t be checked, doesn’t say which parts or how much, and binds no one. It’s a claim.

VERIFICATION

“Check it yourself.”

Provenote confirms how content was made from signals a third party can verify, a watermark, a credential, a signed record, without taking anyone’s word for it.

WHO IT’S FOR

The receiver.

The value is for whoever must trust or comply, a journal, university, platform, or court, and for an honest writer defending against a false accusation.

Just disclosing on a low-stakes personal post? You may not need us, and we’ll say so. Read the FAQ →

Start the record before anyone asks for it.

Join the private beta. Tell us where you feel the cost of getting provenance wrong.

What happens next. We reply within two business days, usually sooner. A short note from a person, one or two questions about what you need to verify, then access if it fits. No automated sequence, and we do not pass your address on.