The first record-keeping system we built for a client was an ordinary database table. It captured everything the AI did, you could search it, and by any normal standard it was fine. Then their compliance officer asked one question that changed how we build everything since: “Can you delete a row?”

The honest answer was yes — of course; deleting things is what databases do. The answer she needed was no. Not “we have a policy against it,” but a flat, built-in no. The gap between a record you promise not to change and one you cannot change is the whole point.

A log, versus proof

Most logging exists so engineers can find bugs. That kind of log is meant to be trimmed and tidied and eventually deleted — nothing wrong with that. But it’s not what a bank or a regulator means by a record. They mean evidence: something you can put in front of an auditor or a board and rely on. And evidence only counts if the people it might embarrass can’t quietly change it afterwards.

The record isn’t a feature of the system. It’s the system’s memory — and it has to be a memory no one can edit.

Why it’s hard

Promising a record is easy; making it genuinely unchangeable takes real work. Each entry is locked to the one before it, so altering anything breaks the chain in a way you can see. It’s written to storage that even we can’t quietly rewrite. And every approval is signed, so an “approved by” can’t be faked later. The reward is that the answer to “can you delete a row?” becomes a confident no.

Who reads it later

The record is read by people who weren’t in the room when it was written — often years later, after the people involved have moved on. An auditor during a review. A regulator after something went wrong. A successor trying to understand a past decision. So it can’t depend on anyone’s goodwill or memory. The test we hold ourselves to: if everyone who built and ran the system left tomorrow, would the record still tell the truth?

Why we start here

We treat the record as the foundation, not an add-on — the thing everything else is written onto. An AI that can act for your business without an unchangeable record of what it did isn’t a tool you can safely rely on; it’s a risk with a good demo. So we build that record first, and everything else sits on top of it.

Next in Notes

Knowing how the AI can go wrong is half the job.

If you can’t describe the ways your AI will make mistakes, you can’t stop them.

Read N° 03 →
Talk to the assistant

This is one of four things the AI always does.

The design agent on the home page sketches a private AI workbench for your business and emails the plan to you.

Talk to the assistant →