Taxonomy / Customary pluralism / Engineering a plurality of conventions in artificial agents

Engineering a plurality of conventions in artificial agents

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.

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

AbsentNot claimed · out of scope

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

Both levers mint conventions inside a single run — a naming game played out over rounds among open-weight agents in NEW-S18, a training loop in NEW-S131 — so nothing has been in use long enough for use itself to confer standing. Neither paper treats a surviving convention as better for having survived: the winning label in the naming game is arbitrary, and NEW-S131's own complaint about self-play is precisely that what it converges on is arbitrary, which is why it adds a penalty to force different answers. Where real human usage does enter, at the point where the trained policy is paired with a human partner, the agent conforms to the partner's existing convention for coordination payoff, not because long practice has earned it authority.

The unwritten constitution

PartialClaimed · partial

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

The uncodified half is fully carried: the orders exist only as emitted labels and first-token preference distributions in NEW-S18 and as weights producing coordinated play in NEW-S131, never as a written rule, and agents coordinate on them without explicit communication. Dicey's other half is missing — binding here means only that a deviator loses coordination reward; nothing sanctions departure, no agent holds another to the norm, and when NEW-S18 rematches partners the label history simply repairs coordination rather than enforcing anything on the defector. Both papers use the word convention and premise that it governs joint behavior without codification, which borrows the tradition's vocabulary, but neither engages the binding-versus-codified distinction that is Dicey's actual point.

Organic amendment

PartialNot claimed

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

In NEW-S18 the content of a convention genuinely is settled from within the population: nobody designates which name wins, agents converge on one through repeated local pairwise exchanges, and when contacts are rewired the retained partner-label history lets the group readjust incrementally rather than be reset — that is the gradual, decentralized adjustment Burke describes. What is authored from outside is everything the approach is actually about: the routing graph decides whether the population unifies or splinters, and in NEW-S131 the entire plurality comes from a cross-play term that pushes each new policy away from the ones already found, plus a mixed-play correction the designer bolts on when that term starts producing bad-faith play. There is no community that holds a convention as its own, so nothing here is the author of its own drift; the drift is a controlled output of wiring and objectives.

Concept Analysis: Newly Introduced

Fragmentation as a control variable

Added

Customary traditions treat plural orders as a historical fact to be understood. Here whether orders stay plural is set by the experimenter's routing graph — the wiring is the knob. That makes plurality something an operator can produce or dissolve, which is a possibility no customary tradition contemplates.

Papers