CapSeal - Fintech onboarding scenario

This is an illustrative scenario, not a customer. CapSeal is pre-launch. Every figure below is modelled from published industry data and stated assumptions, which are set out in full further down so you can disagree with them. We will replace these with named, measured references when we have them, and not before.

The situation

Where the organisation starts

A fintech onboards roughly 240,000 customers a year, with a selfie and document capture in-app.
The liveness vendor analyses the submitted frames. It cannot see whether those frames came from the camera or from a virtual camera driver.
Synthetic identity onboarding is the fastest-growing loss category, and every tightening of the liveness threshold rejects more genuine customers.

Modelled outcome

What the model produces

99.2%

of genuine captures pass without added friction

-64%

modelled synthetic onboarding reaching account opening

8 sec

median added to the onboarding flow

changes required to the existing liveness vendor

The assumptions behind those numbers

240,000 onboarding attempts a year, with a 71% completion rate.
Synthetic identity attempts at 1.4% of attempts, from published fintech onboarding fraud data.
Reduction applies to injected-feed and screen-replay attacks, which the pipeline checks address directly. It does not extend to a deepfake displayed to a real camera in ideal conditions.
Added time is the integrity measurement on device, not network latency.
Existing liveness and document checks continue unchanged and in parallel.

What changes

How the process differs

The pipeline is checked, not just the picture

CapSeal establishes the frame came from a physical sensor, using metadata the operating system attaches that an injected surface must fabricate consistently.

Parallax does the work a classifier cannot

Scene-consistent parallax across a short burst, cross-checked against the device's motion sensors. A screen is flat and a printout is flat.

The platform vouches for the device

Play Integrity and App Attest bound into the manifest. That judgement is Google's and Apple's, not ours, and you can verify it independently.

Genuine customers notice nothing

The measurement happens during the capture the customer is already performing. There is no additional step and no additional prompt.

Honest limits

What this model does not claim

A model is an argument, not evidence. These figures show what follows if the assumptions hold for your organisation, and the point of publishing them in full is so you can check whether they do.

Sealing evidence at capture does not detect every fabrication, does not establish that a truthful-looking document is true, and does not determine intent. The Product Disclosure sets out the boundary precisely.

Test it against your own attack set.

Bring your red-team samples. We would rather find the gap than sell around it.

Book a pilot Check a file first