Contested custom

When people disagree about a claim, this line first works out which raters habitually rate things differently, then keeps only the annotations that both camps independently call helpful. The contest is settled by demanding endorsement across the divide rather than by majority vote: one system selects among notes people already wrote, a later one has a language model draft fresh notes and scores them against a learned stand-in for a divided audience. Because a population's judgments keep moving, the fit is redone as new ratings arrive, and a further proposal makes that drift an explicit variable so the target is re-estimated over time instead of frozen at one measurement.

Instead of writing the custom down, this line stands up a population of model agents that already lives by one and lets conformity, dissent, rebuke, and minority pressure do the work, so the norm in force is whatever the group arrives at. The studies so far run that mechanism and record what it produces: agents publicly go along with norms they privately reject, a single dissenter can break a false consensus, a peer's sanction can teach the norm, and a committed minority can drag a whole population its way, with a response function calibrated on the population's ordinary behavior forecasting how far. Everything here stops at simulation: the group dynamics are genuinely exercised and characterized, but no system yet aligns a deployed model this way.