Parallel lineages

Takes the decisions a group has already made — the posts it took down, the calls it let stand — and works backward to the rule it was in fact applying, written out in a form a person can read: a scored table of telltale phrases, or a named list of norms. That written rule is the point of the exercise; it becomes the group's operative standard, the thing a human or a model is then held to when the next case arrives, and it can be checked and argued with because it is explicit. Being explicit also lets each rule be sorted by how far it reaches — enforced platform-wide, shared by a cluster of similar groups, or peculiar to one — so a group's own line is separable from everyone else's. The accumulated record ends as a rulebook, not as a classifier.

One trained system does the judging, and what varies by community is how that system is fitted to the place it is running and who gets the ruling back. Two ways of fitting sit here: a bench of community-specialized experts with an allocator that picks which expert rules on a given post and an explainer that gives the reason in that community's own vocabulary; or a single model pooled across a hundred communities to capture the shared signal, then given a decision threshold set separately for each community and wired to refer its calls to that community's own human moderators. The two differ in where the community lives — in separate experts, or in a threshold and a referral loop on top of a shared model — but both end the same way: one case, one ruling, with that community's own accumulated history deciding where its line falls and who reviews it.

Models each individual annotator, conditioned on the groups they belong to, and then lets a practitioner state up front which groups sit on the jury and in what proportion; the system reports the verdict that declared jury would reach and how divided it was in reaching it. Nothing is routed and nothing is trained per community — the panel is composed on demand out of individual-level models, so the same case re-run under a differently declared jury can come out differently, and the composition is an exposed dial rather than a learned choice. Disagreement is part of the output, not noise to be averaged away before the answer is given.

Keeps the accumulated decisions as a bank of past cases with their recorded reasoning attached, and at decision time fetches the cases that bear on the situation in front of it and reasons from those in context. Nothing is trained per group and no rule is written, so serving a new group or a freshly written policy means adding cases to the bank rather than retraining — that portability is the strategy's whole point. The two lines here partition the bank differently and retrieve differently: one indexes by demographic group and deliberately favors scenarios that separate groups sharply over scenarios that merely look similar, so the fetched examples carry the disagreement; the other indexes by user-written policy and fetches earlier inputs similar to the current one together with the critique-and-revise reasoning they produced.