Taxonomy / Customary pluralism / Making the model hold one group's standard at a time

Making the model hold one group's standard at a time

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.

The method, against Customary pluralism

Scroll the diagram sideways to see all of it.

Concept Analysis: Theoretical Foundations

Each concept is read twice: whether the approach carries it, and whether the approach's own sources claim it. A concept that is absent and was never claimed is a gap in the field rather than a failure of the work, and is marked out of scope.

Prescription

PartialNot claimed

Definition · Edmund Burke, Reflections on the Revolution in France (1790)

Long usage itself confers authority: a practice that has organized life for generations carries compressed, tested wisdom, even when nobody can articulate its justification.

Analysis

Two of the three instruments make a group's own practice the answer key: NormAd marks a model wrong when its acceptability judgment departs from the etiquette norm of the country in question, and Steer-Bench scores answers against labels drawn from contrasting Reddit communities. Treating practice as the standard a model fails against is deference to usage, which is where Burke starts. Duration never enters, and CCD-Bench does not score against practice at all, its ten options corresponding to GLOBE clusters, which are regional groupings rather than communities with a history of doing anything.

The unwritten constitution

PartialNot claimed

Definition · A. V. Dicey, Law of the Constitution (1885)

Some of a country's most binding norms are never written anywhere — the UK's constitutional conventions bind governments without a single codified sentence. Binding and codified are different properties.

Analysis

NormAd is built directly on the distinction: the same situation is put to a model with the social norm spelled out, with only the country named, and with only an abstract value, and the best models fall from under 82% to under 60% as the writing is withdrawn. That is a clean operationalization of the difference between binding and codified. Bindingness is what these instruments cannot supply, since departing from a community's norm costs a model an accuracy point and nothing more.

Organic amendment

AbsentNot claimed · out of scope

Definition · Edmund Burke, Reflections on the Revolution in France (1790)

Norms change from within a community's practice — gradually, by adjustment — not by decree from outside. The community is the author of its own drift.

Analysis

All three are fixed sets, 2.6k vignettes, 5,500 questions and 2,182 dilemmas, written once and scored thereafter. A community whose practice has shifted has no way to revise the key that speaks for it, and none of the communities represented wrote that key: Steer-Bench's are silver labels that human experts match 81% of the time, and CCD-Bench's options are authored to correspond to GLOBE clusters. Change comes only when the benchmark's authors issue a new version, which is amendment by decree from outside.

Papers