How matching works

Joan talks with you through four short conversations: how you handle closeness and distance, how you give and receive care, your dealbreakers, and how you handle stress, conflict, and repair. The questions are open. You talk, she listens, and a separate analysis step reads the transcript afterward. Nothing is scored mid-sentence.

From your attachment and emotional answers, a language model rates eight signals: attachment anxiety, attachment avoidance, communication openness, conflict response, comfort with intimacy, independence, trust trajectory, and caregiving style. Each rating carries a confidence score, and thin evidence is deliberately capped: if you barely touched a topic, the system is not allowed to pretend it knows you.

You will never see yourself reduced to a clinical label. Our own clinical review was blunt about this: these are bands and narratives, not diagnoses, and they are treated as priors, not gates.

The closeness half runs on one peer reviewed anchor: a 2012 meta-analysis (Li and Chan) finding that attachment anxiety and avoidance both predict lower relationship quality, with avoidance the heavier signal. We weight it the same way: avoidance 60, anxiety 40. Two people's combined load sets the base score, and the other six signals adjust it up or down by at most 12 points.

The final number blends three parts: 60 percent that closeness score, 20 percent how you give and receive care, 20 percent dealbreaker fit. Those weights are honest guesses by us, not calibrated constants. We will tune them against real outcomes when we have real outcomes, and the council question about the blend is live right now.

When you tell Joan your dealbreakers, we extract positions on five topics: kids, religion, substances, relocation, and money. A position is only recorded when the model is at least 70 percent confident and can quote your exact words back. Below that, we store unstated and never guess.

Only three topics can actually hide a person: kids, substances, and religion, and only when one person's stated hard requirement directly opposes the other's stated position. Relocation and money never hide anyone; they are context, not filters. A near miss (like occasional use against a firm substance-free ask) does not hide anyone either; it shows up as worth discussing, and it withholds the no-conflicts checkmark. When you both stated the same position, we say so.

• The 70 bar: we only show pairs scoring 70 or higher. Council question live.
• The blend: 60/20/20 across closeness, care, dealbreakers. Council question live.
• The radius: 25 miles by default, a hard wall today. Council question live.
• The confidence floor: 70 percent, never guess. Council question live.

Every decision lands in the ledger with before and after values.

Vote on the numbers.

We do not diagnose you. We do not show anyone your raw scores or labels. We do not learn from your swipes or track you around the web. We do not claim these numbers are scientifically proven. One anchor is peer reviewed; the rest is careful design that our own experts challenged hard before it shipped, and the honest word for the weights is chosen, not discovered. That is exactly why the council exists: the choices are real, so you should get a say in them.