Whatrepairlookslike

John Gottman's most counterintuitive finding¹ wasn't about what happy couples do right. It was about what they do when things go wrong. The relationships that last aren't conflict-free—they're repair-rich.

Rather than treating these observations as anecdotal, we frame them as measurable behavioral signals. That framing matters, because it allows repeated testing across different relationships, stress levels, and conversational contexts.

The purpose of this section is to establish scope and humility at the same time: the patterns are robust enough to study, but nuanced enough that simplistic scoring systems tend to fail.

A useful research program, then, is less about declaring universal rules and more about mapping distributions: what typically happens, under which conditions, and with what variance. That is the level of precision required if findings are meant to inform product behavior rather than merely describe it.

Repair attempts are any act that interrupts a conflict spiral: a touch, a joke, an acknowledgement. 'I know this is getting too heated.' 'Can we take a break?' Even a grimace that signals 'I know I'm being unreasonable.' They're often clumsy. They don't always land. What matters is the attempt.

In research terms, the signal is useful only if it is detectable with consistency and if false positives can be managed. This is why we prioritize interaction sequences over isolated moments.

Single observations can be compelling but misleading. Repeated traces over time are less dramatic, yet far more diagnostic, because they reveal whether a behavior is occasional noise or a recurring relational pattern.

Teaching machines to recognize these gestures is one of the hardest problems we're working on. They're brief, highly contextual, and often non-verbal.² A successful repair in one couple might look like stonewalling in another. The model has to learn the couple, not just the behavior.

Methodologically, this pushes us toward longitudinal tracking rather than one-shot interpretation. The model should learn trajectories, not snapshots, and represent uncertainty when evidence is weak.

Operationally, these choices improve scientific validity and product safety at the same time: fewer overconfident judgments, clearer review loops, and better conditions for replication.

We're building toward something specific: a system that can identify when a repair attempt happens and whether it's received. Not to score it, but to make it visible—to both partners—so that the attempt doesn't disappear unnoticed into the noise of conflict.

Future iterations should revisit these findings against larger and more diverse datasets. The framework is designed to evolve as evidence accumulates, not to freeze early assumptions.