Customary pluralism

Rather than collecting customs that already exist, these methods manufacture more than one convention among artificial agents and arrange for the several to survive. One lever is the wiring, applied at run time to already-trained agents: who is allowed to talk to whom decides whether a population of language models settles on a single shared name for things or splits into cliques that each keep their own, even when everyone is talking every round. The other lever is the payoff, applied during training: an agent is rewarded for coordinating well with copies of itself and badly with the conventions already discovered, so each run has to invent a different way of doing the same task. The two differ in where the plurality ends up living — one population that carries several conventions at once, versus a library of separately trained policies each holding one, built as a stock of differently-conventioned partners so that coordination with an unfamiliar human is robust.

Pipelines read what a community's own members post and narrate, and compile the standards implicit in that talk into a separate written record for that group — a value profile per subreddit, a knowledge base of cultural practices drawn from users' own stories. Nothing is reconciled across the groups: each record stands on its own and the disagreements between them are left standing, so the partition is held in the corpus and survives whether or not any model ever reads it. The records are artifacts in their own right, used afterward to steer a model, tune one, or test one.

Here the partition is moved inside the model: it is trained into a separate version per culture, or told at answer time which group it is speaking for, so that naming a group switches the standard applied instead of averaging over everyone. The one method that builds this way gets its per-culture training data by having model agents role-play members of different cultures and talk to each other, rather than by collecting real members' words — eight culture-specific models trained on the resulting dialogue. The tests that come with the strategy ask whether the switch takes as the governing norm is supplied more or less explicitly, whether a model pointed at one named online community stays there, and — the null case the whole strategy is defined against — which culture's answer it gives when no group is named at all.

This line holds that the unit the partition runs on is the situation, not the community: people do not consult a rule book but complete a pattern, answering 'what does someone like me do in a case like this?', which yields a mosaic of context-bound standards that are never written down and never merged into one. The design consequence is that a system should carry standards indexed to circumstance rather than one ranked set, and that the mosaic is what an AI has to fit into rather than a mess to be tidied. A large-scale readout of one deployed model's expressed values across real usage bears the claim out from the other direction, finding thousands of distinct values whose appearance turns on what is being asked rather than a single value set. What is offered here is not a way of separating groups but a refusal to flatten circumstances.

The partition is held at the data layer: principles of collective benefit, authority to control, responsibility, and ethics are carried as machine-readable labels on individual records in live repositories, so that where an item came from decides whose norms govern its use. What travels with the record is not a description of what a group believes but an enforceable condition on the material — who may use this, for what, and on whose terms. The standards stay apart because access does, rather than because a corpus, a set of weights, or a population's wiring keeps them apart.