The Matchmaker Algorithm

Last November I had an idea. What if you could predict relapse before the person experiencing it had conscious access to the fact that it was coming?

Not from a check-in. Not from a survey. From the ambient digital behavior they were already generating — the stuff they weren’t trying to communicate to anyone. Sleep patterns through phone movement. Typing speed at different hours. App opens at 2am. Location entropy. Music choices. The kinds of signals that don’t require you to do anything except live your life with a smartphone in your pocket.

The Study

Two cohorts. One group: people who have voluntarily entered substance abuse treatment. They know what this study is about. They consent to the quantitative UA — random, not scheduled, truly random — as the ground truth. The other group: people who’ve agreed to let the same ambient telemetry run while they simply live their lives, no treatment context. Non-clinical. Control group by disposition, not assignment.

No PII collected. No identities associated with the data. You’re not watching people — you’re watching patterns.

Then you run the regressions. Quant UA result as the dependent variable. Everything else as independent. You’re looking for p-values. You’re looking for which ambient signals, in which combinations, have predictive power across a population that the individual couldn’t have told you themselves.

If it works — and I think it does — you have something.

The Intervention

What do you do with it?

This is the part most people get wrong. The obvious answer is: flag the person. Tell a clinician. Insert a human into the loop. Create a case. That is exactly the wrong answer, and here’s why: the moment a person knows they’ve been identified as a relapse risk, you’ve broken the thing that makes passive data useful. You’ve turned an ambient signal into a label, and now they’re managing the label instead of their behavior.

The system I want to build doesn’t do that. Ever.

What it does instead: when the model says someone’s risk vector is elevated, the system randomly connects them with another user for a few minutes. A chat. Completely framed as organic — the platform matched two people who might have something to say to each other. No reason given. No flag visible to either party. No human anywhere in that loop who knows why the connection happened.

A user is never identified as a relapse risk to any human being. Not a clinician. Not a researcher. Not a support person. Not their family. The math saw something, the system responded, and both participants experienced a moment of human connection that they may or may not remember a week later.

That’s the intervention. An interruption that doesn’t announce itself.

The Part That’s Actually About Dating

Once you see this architecture, you see it everywhere.

The thing I’m describing is exactly what the best human matchmakers do — the ones who actually work. Not apps. Not algorithms that surface mutual hobbies and compatible income brackets. The woman in your social circle who just knows. You don’t know her methodology. She doesn’t tell you. She doesn’t ask you to fill out a form. She watches how you move through a room, what you light up around, who you’re drawn to before you’re conscious of being drawn. And then one day she says, “you should meet this person,” and somehow she’s right.

The ambient data is the footprint. The regression is her intuition formalized. The random connection is the introduction she engineers without explaining herself.

The app version of this isn’t swiping. It’s not surfacing profiles. It’s the system noticing something about where you are in your life and quietly arranging a collision. You experience it as coincidence. The math knows better.

The Reason This Matters Beyond Dating and Addiction

I want to be direct about this part, because the whole architecture I’ve described — passive data collection, behavioral pattern extraction, predictive regression, targeted intervention — is also the exact blueprint for some of the worst systems humans have ever built.

Pre-crime. Thought police. Social credit. Predictive policing. These are not dystopian fantasies. They are present-tense realities in functioning democracies, built on exactly this mathematics applied at state scale with punitive intent. The same p-value that predicts relapse predicts recidivism. The same ambient signal that connects two lonely people can mark one of them for surveillance. The model doesn’t know the difference. The math is neutral. The architecture is neutral.

What isn’t neutral — what has never been neutral — is who controls the output and what they’re authorized to do with it.

That’s the actual ethical question. Not “should this math exist.” It exists. It will continue to exist. The question is: does the system that holds it have authority to act punitively on what it finds? Is the person ever labeled? Is the action ever coercive? Is there a human anywhere in the loop who gets to decide what a flag means for your life?

My intervention says no. Categorically. The system connects people. That’s all it does. No one is identified. No one is punished. No record follows anyone.

That boundary — the commitment that the model’s output is only ever used to facilitate connection and never to assign status, risk, or consequence — is not a feature. It’s the entire ethical distinction between what I’m describing and what a surveillance state does with the same math.

We need to build this shit right, or we shouldn’t build it at all. Because the version where someone makes the wrong architectural decision — where a researcher decides it’s more efficient to flag high-risk users to a care coordinator, where a platform decides a risk score is more monetizable than a random connection — that version isn’t helpful. It’s just another surveillance system with better PR.

The same mathematics. Completely different intent. And intent, encoded in architecture, is the only thing that separates them.

This connects to the governance framework I’ve been building in AIGCSEP — specifically what it means for an AI system to act without surfacing its reasoning chain to any human actor. The scoped intervention, the no-identity commitment, the vault pattern: those are governance decisions, not product features. They’re the answer to how you apply predictive capability without accidentally building a surveillance apparatus.

The matchmaker doesn’t explain herself. But she also doesn’t mark anyone.

That distinction is the whole game.

— J.P. Howlett