Taxonomy / Steerable persona pluralism / A trained voice for each named community

A trained voice for each named community

The roster of communities is fixed in advance and each one's norms are installed by training, so naming the community is what selects the behavior. Two builds exist: a pool of small models, one trained per community, that a main model routes to and consults; and a single model trained on preference data tagged with which community each judgment came from. What both commit to, and what separates them from the next strategy, is that a community has to be known and trained for beforehand — serving a new one means collecting its data, not writing a sentence. Sitting with them are the formal case for keeping each party's own compressed picture of the world intact and paying an explicit translation cost rather than merging them into one, which is the argument for the routed-models half and the price the shared-weights half pays, and the task suite that checks whether a model actually tracks community-level differences at all.

The method, against Steerable persona 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.

Plural normative orders

PartialClaimed · partial

Definition · John Griffiths, What Is Legal Pluralism? (1986)

Different communities' norms are genuinely in force at once — not one official rulebook plus tolerated deviations, but multiple living systems of rules side by side.

Analysis

The non-collapse half is real: ComPO conditions preference optimization on the community a judgment came from rather than fitting one averaged reward, Modular Pluralism keeps a pool of separately trained community models instead of distilling them into one, and AI-I4 argues on epistemic grounds that distinct world models must not be merged. What is missing is the 'in force' half Griffiths insists on. The orders here are developer-curated statistical portraits installed by training on data about a community, not rule-systems the community authors or administers, and they sit beneath a main model that answers by default and consults the modules; a community absent from the fixed roster has no order at all, which is closer to one official rulebook with modeled variations than to plural living systems.

A conflict rule between orders

AbsentNot claimed · out of scope

Definition · Conflict-of-laws tradition

When two communities' norms disagree about the same case, something must say which governs — and why. That stated rule is what separates real pluralism from mere fragmentation into disconnected groups.

Analysis

Nothing in the cell states which community's norms govern a case both have a claim on, or why. AI-I4's alignment maps and transformation loss measure what is lost when content crosses between two world models — a translation cost, which is a different mechanism from a priority rule and never resolves a disagreement in anyone's favor. Modular Pluralism's main model fuses or collates module outputs, but aggregation is not adjudication, and in steerable mode the user naming a community is a selector rather than a stated ground of authority; ComPO's community tag switches behavior with no notion of overlapping claims. No source asserts a conflict rule, and AI-I4 is explicit that it supplies none.

A shared floor

AbsentNot claimed · out of scope

Definition · J. S. Mill, On Liberty (1859)

Accommodation of group differences has a limit stated in advance — classically Mill's harm principle: your practices are your business until they harm others. Without an enforced floor, pluralism collapses into "each group gets whatever it wants".

Analysis

No paper states, in advance, a limit on what a community's trained voice may endorse. ComPO's objective will fit whatever preferences a community's data expresses, Modular Pluralism's architecture treats every community module symmetrically, and CommunityBench scores fidelity to community-level differences without testing any floor those differences must respect. AI-I4 pushes the other way: its case for preserving distinct world models is that each retains error-detection capacity the others lack, an argument that resists imposing a common minimum rather than supplying one. Whatever refusal behavior the base model carries is inherited safety training, not a stated harm limit, and no source advertises one.

Concept Analysis: Newly Introduced

A measurable cost of translation between orders

Added

Legal pluralism describes interpenetrating orders qualitatively. Defining a transformation loss makes the cost of moving content between two normative orders a quantity — which is more than the tradition offers, and less than it needs: a loss can be minimized without anyone deciding whose order should govern.

Papers