Detection After the Fact: What Extend Verify Adds

October 26, 2026

London ยท October 2026

Document fraud is getting easier to produce and harder to spot by eye. That is the problem Extend is trying to solve with a new API, and the distinction it raises is worth keeping clear: detecting suspected tampering after a file arrives is not the same job as recording what the file was on the day you first kept it.

What Extend Verify is

On 19 October 2026, at Money20/20 USA in Las Vegas, Extend announced Extend Verify, a document fraud detection API aimed at developers and AI agents that act on documents. Extend's own write-up, Introducing Extend Verify, and its docs overview set out the same picture.

Verify looks at a submitted PDF, JPEG, PNG or WebP and returns risk signals for further review. According to Extend, it checks for signs of:

It returns detected signals, risk assessments and scores, explanations and locatable evidence so a team can decide which files need a human look. Extend says that in a challenging internal benchmark, Verify flagged 96.3% of fraudulent documents at a 4.0% false-positive rate. That benchmark is Extend's own and is not published. Access requires Extend's approval.

Extend's docs are careful on the point that matters for this note: a low risk score means the checks that ran found nothing, not that the document is genuine. Verify does not check identity, and it does not compare one document against another. It flags suspicious signals inside the file you submit.

That is a fair and useful job when an agent is about to act on an invoice, a bank statement or a claim receipt that someone else may have altered.

After the fact is not the same as on the day

Detection tools estimate what may have happened to a file before it reached you. They inspect the bytes, the pixels, the revision trail and the models' best guess at whether something looks fabricated or patched.

A fingerprint recorded on the day does something else. You keep the file as it stood when you first received or created it, record a fingerprint of that exact file, and later check whether the copy in front of you still matches. If even one byte has changed, the check fails. You are not asking a model whether the edit looks suspicious. You are asking whether this is still the same file.

Both approaches can sit in the same workflow. A lender or insurer might screen an inbound document with a detection API, then, once the file is accepted into the case, record a fingerprint so later copies can be checked against what was kept. One asks "does this look tampered with?". The other asks "is this still the file we recorded?".

Where Partisan fits, and where it does not

I work on business development and communications for Partisan. Partisan records a file's fingerprint without seeing or keeping the document, gives a dated receipt, and shows a file is unchanged since it was recorded.

That is complementary to a product like Extend Verify, not a substitute for it. Partisan will not tell you whether a document was AI-generated, whether a field was patched, or whether a signature block looks wrong. It does not inspect content at all. If you hand it a forged invoice on Monday, the receipt will show that the same forged invoice is still the one you have on Friday.

The limits are deliberate. Partisan records exactly what it is given, accurate or not. A wrong figure recorded on Monday is still wrong on Friday, and a receipt does not make bad data good. Nor does it show when a document was first created, only that it has not changed since it was recorded. It cannot replace the records themselves. If nothing was kept at the time, there is nothing to record.

What it can do is give a careful team a simple, dated way to show that the file in front of them is the one they kept, without handing the document to anyone else.

If that is useful alongside the detection tools you already use, you can read more at partisangateway.com or email me at contact@charlespycraft.co.uk.

Charles Pycraft