EXHIBIT / FROM ROOMS 6 AND 7 OF THE WALK

Sabit

I treat month-end close as a locked room, the AI can advise all it wants however it is never allowed to touch the math.

Picture month-end close as a locked room. The numbers go in, every account and every journal line, and the door does not open until debits equal credits and every human override is logged with a name and a timestamp. There are no exceptions for a deadline and none for an executive either, since the room certifies math and being senior does not get you a key.

There is an AI in the room, but it sits at the visitor’s desk. It can read everything, notice the gap, suggest the fix, explain the variance. It cannot type. The moment a number needs to change, a human changes it, and the change itself becomes part of the record.

The whole product is enforcement. Anyone can build software that finds the imbalance. I built Sabit so that nothing, human or machine, can call the books closed until they provably are.

Below is the smallest possible version of that room. Close it yourself.

Loading the trial balance…

With JavaScript off, the room stays locked on paper. The trial balance is out by 600.00 on the credit side. The advisor can point at the missing utilities accrual but cannot enter it, and when a human supplies the number and the Truth Gate re-checks, the books snap into balance and a CERTIFIED result prints.

What it is for

Costs today: A finance team that wants AI at close, and an auditor who needs certainty the AI changed nothing. Both are right, and today they cancel each other out.
What changes: The AI reads and advises at full speed, the math stays enforced, and the auditor gets a record with a name on every change.
Known in 90 days: Close cycle time against the same books, and the count of overrides that reached the record without a name. The second number should be zero.

Why I set it down

Sabit was also a product bet. Partway through the build, Anthropic released finance templates that overlapped it, and in my humble opinion a bet is only worth holding while you would place it again today. That morning I would not have.

So I retired the bet the same week and kept the architecture, the enforcement pattern and the Truth Gate, under everything I have built since. What would reopen it is a regulated buyer who needs enforcement the templates do not give, with their own auditors in the room.

It took a year to build and a week to retire, and if you are reading this to decide whether to hire me, the week is the part I would point you at.

Go deeper into the thinking

The part of Sabit’s README that holds up best is the list of things it refuses to be. Every “not this” below is a market someone would have chased, and I walked past them on purpose.

What this is not

NOT THISNot an ERP. It certifies the close and leaves running the business to whatever already does.
NOT THISNot forecasting. It proves what happened, and stops there.
NOT THISNot payments. Money never moves through it, evidence is the only thing that does.

This is how I would make an AI program audit-ready: put the AI where it reads and advises, and leave a record a regulator can walk. In close terms that maps to SOX 404 and ICFR controls, where the question is always who could change a number and whether the change carries a name. The record is append-only, so a correction is a new line with a name on it, never a quiet edit of the old one.

Source lives here (repo CPACFA); read it on GitHub →

Next: Tanaka →