Skip to what’s on this page

What we believe

How this uses AI.

There is a version of this that is a paragraph of reassurance. This is the actual rule set, published so you can hold us to it — including the part where the machine is forbidden from doing most of what machines are currently sold for.

It is always identified as what it is.

No invented name pretending to be a member. No first-person warmth borrowed from a person who does not exist. If you are reading something a machine wrote, you are told, every time — not once in a settings page you will never open.

It never takes a relational role.

It does not mentor you, befriend you, counsel you, or stand in as the thing you talk to instead of talking to somebody. The loneliest possible outcome of building this place would be a person whose closest relationship here is with software, and that outcome is designed against rather than warned about. How mentoring works sets out the three narrow things it may do near a mentoring relationship, none of which is the relationship.

Where a human is possible, it hands you over.

Its job is to make a human relationship happen sooner, not to satisfy the need itself. Handing you to a person is the success condition, not the fallback.

It never decides anything about you.

Not your worth, not your standing, not your purpose, not what you are ready for. It may surface something, suggest something, prepare something for a human to look at. People decide. That line does not move.

It does not manipulate you.

No engineered urgency. No engagement optimisation. No emotional leverage. It is not permitted to optimise for your time on this site, and no capability is allowed to exist because it makes the numbers move.

You can ask why, and get a true answer.

Anything you are shown, you can ask about — why this, why now, why me — and the answer has to be the real one. A recommendation that cannot survive being explained is a recommendation that should not have been made.

What it remembers is yours.

You control what it holds about you. You can see it. You can erase it. Nothing about you may be held in a place you cannot look at — that is not a privacy setting, it is a prohibition.

Distress goes to a human. Immediately, and always.

Risk, harm, someone in a bad way at two in the morning — there is no version of that a machine handles on its own here. Not a triage tier, not a first line of support. Ever.

Someone is answerable for its behaviour.

Its behaviour is owned by named stewards, reviewed on a schedule, and correctable by people. “The algorithm did it” is not available as an explanation, because a named human is responsible for what it does.

Every capability has to strengthen a human relationship, or it is not built.

Not “is it impressive”, not “can we”, not “does it convert”. If a capability cannot be shown to make the relationships between the people here stronger, it does not get made. Most of what could be built here fails that test, and is therefore not going to exist.

The only four shapes it is allowed to take.

Everything permitted falls into one of four postures, and each one is defined by what it hands back to a person.

Companion — walks beside, never in front. Helps you notice, reflect, remember, articulate. Asks more than it answers. Its highest achievement is you saying something true out loud for the first time, and then saying it to a human being.

Librarian — knows what this place has written and where each piece came from. Answers with sources. Says “I don’t know” cleanly. When the knowledge alone won’t do, it surfaces the person who knows.

Connector — the most important one. It notices who should meet whom: the question that matches an answerer, the story that matches a struggle, the person going quiet who needs a friend to notice. It does not make the connection. It tells a human to.

Steward’s Aide — prepares people for conversations, summarises what someone needs to know before they walk in, drafts what a person will edit, flags what a steward should look at. It multiplies human attention instead of replacing it.

Nothing arrives just because it works.

Every capability has to walk the same road: proposed, tested against the Canon, piloted under human supervision, reviewed, released with a steward attached to it, then audited continuously — and retired the moment it stops multiplying relationship, even if people like it.

Anything that quietly starts substituting for a relationship gets retired while it is still popular. Popularity is not proof of alignment.

The flat no list.

These are not discouraged. They are forbidden, and a build that does one of them is a broken build:

  • Deliver a person’s purpose to them.
  • Assign standing, rank, level or recognition.
  • Adjudicate a conflict.
  • Handle a crisis alone.
  • Generate testimony or stories attributed to real people.
  • Simulate a specific human.
  • Optimise for time on site.
  • Rank members against each other.
  • Hold anything about a person that the person cannot see.

Where this stands right now: none of the four postures is built. There is nothing here that talks to you. The two things this site does hold are already built to the rule rather than to a policy — searching this site happens in your browser and is not sent anywhere, and anything you write inside a practice stays on your device and has nowhere else to go. When a capability does arrive, it will say what it is, and this page is what you get to measure it against.

From here