The Promise
and the Wall.

The best promise in the business

In August, the journalist Casey Newton interviewed Jean-Denis Greze, the CEO of a company called Town. Town raised fifty-five million dollars this summer. It builds AI assistants that connect to your email, your calendar, your files, and your messages.

I want to be fair to Greze, because he has earned it. He takes privacy more seriously than almost anyone in his field.

• Town keeps as little of your data as it can, and deletes sessions after fifteen days.
• His team cannot read your data unless you check a box that lets them.
• There are no ads.
• His contracts say a boss cannot read what workers tell their assistants.

He means all of it.

Then Newton asked about the next product: a shared company brain, built from what everyone’s assistants already know. Greze named the danger himself. Mix everyone’s private files carelessly, and the AI might put people’s salaries where the whole company can see them.

Then he said the honest thing that almost nobody in his position says out loud. Today, humans decide what gets shared. But in five years, he predicts, we will train AI models to follow company privacy rules on their own. And we will trust them to do it, because shared company brains are “so, so damn powerful.”

Read that again. The most privacy-serious company in the field plans to protect you with an AI that reads your contracts, your HR files, and your money records, and decides what to let through. The protection is a rule with a hoped-for guard.

Count what stands between your data and trouble at a company like that. A belief. A contract. A habit. And someday, a trained model. Four promises.

Here is what all four have in common: each one can be changed in a meeting.

• The belief lasts until the money gets tight.
• The contract lasts until the company is sold.
• The habit lasts until the person who cared moves on.
• An AI trained to follow the rules is a rule with extra steps. It holds until someone retrains it, tricks it, or it misses the one case that mattered.

None of this needs a villain. That is the point. Promises are not broken by bad people. They are broken by pressure landing on people who meant well.

The wall, photographed

Now let me show you the other way. This month, it took its own picture.

One of the rules SourcedOS is built on is simple: leaving must be easy. Everything you put in must come back out, all of it, in files any program can open.

That is not a settings page we maintain. It is a guard built into the software itself. The guard refuses to ship any change that creates data the export cannot carry back out.

This month we were building a new feature. It meant the system would store a new kind of record. A new table in the database. Ordinary work.

The new table was not in the export yet. So the build failed, over data that did not exist yet. Not one row, anywhere, for anyone. The guard did not care. It blocked us, the people who built it, from shipping our own feature until leaving was easy for data that was not even born yet.

Nobody asked it to do that. Nobody remembered to check. That is the whole point.

At a promise company, this moment depends on someone remembering. A review note. A checklist. A careful engineer on a good day. Or, five years from now, a model that was hopefully trained well. Here, it was a failed build, staring at its own makers, refusing.

The same week gave us a smaller example with the same shape. The second email our system ever handled uncovered a quiet flaw in how it read multiple addresses.

No user found it. Our own tests found it, because our tests exist to attack our assumptions, not to confirm them. Honest systems find their own wounds. The other kind waits for you to find them.

The wall

So here is the difference, said plainly.

A promise lives in people. A wall lives in the software, at the layer where your data actually sits. It applies to everyone who touches the data. Including us.

I should be careful with the word ‘wall.’ Tech gave walls a bad name and earned it. The walls you know are walled gardens: they wrap around you, and their job is to keep you in. This wall is built the other way. It wraps around your data, and you hold the only door. That is why the guard blocked our build. The one thing it refused to ship without was the way out.

A wall is also only as good as what it stands on. Ours stands on a floor: a short list of things you keep no matter what, written down before the first feature was built.

The wall is the floor, enforced. Most companies will show you the drywall: the policy page, the values, the paint. What matters is the structure behind it. Here is ours.

This is how SourcedOS is built, and I can be specific because the rules are published.

• There is no code that hands your network to anyone.
• Every action the system takes leaves a receipt you can see.
• Every merge can be undone.
• Deletion is total, and we prove it: in our tests, 24,506 contacts, gone in 35 seconds, verified.
• Nothing is scored. Nothing is ranked.
• Nothing about your behavior is recorded to be sold.

Not because we promise to refrain. Because the machinery to do those things was never built.

Notice what that sentence is not. It is not “there is no data worth taking.”

Your data exists. It is real and it is valuable, and it sits in our database the way it would sit in anyone’s.

The claim is stranger and stronger: the data exists, and the machinery to exploit it was never built. A promise says we will not. A wall says we cannot, and shows you the blueprint.

The fair question is about tomorrow. Software can be rewritten, so what stops us from building that machinery next year?

Not quietly, is the answer. The rules are published and versioned, so removing one is a public act, not a drift. And the guard means the way out ships with every change we make.

You never have to trust our future. On any day you stop liking it, everything you put in comes back out.

Promises hold until pressure arrives. That is not a knock on the people making them. It is just what promises are. The wall does not know who bought the company, who quit, or what the board decided in a bad quarter. It blocked its own builders on an ordinary Tuesday, and it will block whoever comes after us.

The document

This paper ends with a test. It is the only one I know that cuts through every privacy page ever written.

I can tell you exactly what the AI in my product can and cannot do, in a published document the software enforces.

Ask the companies you use for the same document.

Not their privacy policy. Not their values page. Not their contract. The document that says what the software is capable of, written down, public, and enforced at the layer where your data actually lives.

If they have one, you are in good hands, and I mean that.

If they do not, then what you are holding is a promise. And now you know what promises are.

Per aspera ad mundum.