Aprentiz home Apply for a founding research place
Menu

Built to be checked.

For your IT provider and the person who signs: what the system does, where it runs, and how we test it.

The FCA, 19 October 2026 (section 4.2, areas for improvement): “Overreliance on untested vendor assurances.” Read the FCA’s review

Built and working.

As at 24 October 2026. Only what is built and tested is listed here.

  • Every decision signed with Ed25519, chained to the one before and tamper-evident
  • Won’t conclude without evidence, and pressure can’t change a signed decision
  • Gaps shown as gaps, never reported as covered
  • Every cited sentence checked against the exact source text
  • Rule versions kept with their dates, from authorised activation onwards
  • Your firm’s perimeter from the FCA Register, by FRN, in the founding programme
  • Private open-weight models, with no fallback to an outside AI service
  • Won’t start if it can reach an outside address
  • Named sign-in with passkeys
  • Encrypted checkpoints, with recovery rehearsed in our R&D cloud
  • Every software component listed and scanned for known vulnerabilities

Check our film’s signature yourself.

The record in our film is signed with Ed25519 when the film is made. Here are the public key, the exact record and the signature, so your IT provider can repeat the check without us. They belong to the film with sha256 6bbb396e7ac68d28….

Download the key, record and signature (JSON)

  1. Save record as a file, exactly as given, with no newline at the end.
  2. Save signature_hex as 64 raw bytes, and the public key as DER: 302a300506032b6570032100 followed by public_key_hex.
  3. Run openssl pkeyutl -verify -pubin -inkey key.der -keyform DER -rawin -in record -sigfile signature. It verifies. Do the same with changed_record, the record with one byte changed. It fails.

Where your work runs.

A private environment for your firm alone, on AWS in London (eu-west-2). During the founding programme, we set it up and run it for you.

  • Example clients

    The founding programme works on example clients, not real client files.

  • Never pooled

    Your work is never pooled with another firm’s.

  • Shared R&D

    Partners share only what we learn about public rules and model behaviour, never anything that names a firm or describes its data or methods.

Which AI models run.

Private, open-weight models that run inside the environment, with no fallback to an outside AI service.

If the private model isn’t available
Aprentiz stops. It doesn’t send your work elsewhere.
If it can reach outside addresses
It won’t start. It checks this at every start-up.
If the model changes
A signed-off decision isn’t reopened without new evidence.

How decisions are signed and checked.

A record of who decided, on what evidence, under which rule, sealed so that a later change would show.

“Some firms’ decisions on AI tools were poorly documented with limited audit trails or supporting rationale.”

FCA, 19 October 2026. AI in financial services: good practice and areas for improvement, §3.3.
  1. A named person in your firm makes the decision and signs it off. The AI never approves anything.
  2. Aprentiz records who approved it, and writes the decision to an append-only log.
  3. The record is signed with an Ed25519 digital signature and chained to the one before. Change one byte and the chain breaks.
  4. New evidence creates a new record. It never rewrites the old one.
  5. Each evidence pack carries the steps to check it. Our standalone verifier checks the signatures without contacting us.

How we build and test.

For your IT provider and your consultant: what we check, and how.

We try to break it before you rely on it.
Pressure, repeated asking and a different model: identical inputs can’t flip a decision.
Every citation is checked against the source.
A quote that changes the meaning, or adds a “not”, fails.
Every decision is signed and sealed.
One changed byte is caught.
It won’t start if data could leave.
We also test that an attempt to reach an outside service is blocked and recorded.
Recovery is rehearsed, not assumed.
Restores are run and checked, in our R&D cloud environment.
We test our tests.
We deliberately break the code to check that our tests notice.

Ask the hard questions. We’ll answer them plainly.

Write to us