Who adjudicates: a Game Master, a motivation function, or nobody

Every simulation of people contains a moment where someone has to make a ruling. An agent attempts something, to buy at this price, to enter this room, to persuade this customer. Some mechanism decides whether the attempt succeeds and what the world looks like afterwards. That mechanism is the real design decision in any agent system, well ahead of which model plays the people. Two papers, 16 years apart, answer it in opposite directions. Concordia, DeepMind’s 2023 library for generative agent-based modelling, puts the ruling outside the agent, in a component it names after the referee of a tabletop role-playing game. A 2007 consumer simulation in the Journal of Business Research put the ruling inside the agent, as an equation. The placements are opposite. The property they share is the one that matters, since in both designs a reviewer can point at the rule and read it.

I keep returning to this pair because the question they answer has become the question I ask of every simulation pitch that crosses my desk. Who adjudicates? “The model decides” is not an answer.

The referee outside the agent

Concordia borrows its central abstraction from Dungeons & Dragons. In the tabletop original, players say what their characters attempt. A human called the Game Master rules on what actually happens. Concordia implements exactly that split in code. Agents consume observations and produce action attempts, plain English descriptions of what they want to do. The Game Master consumes those attempts and produces event statements, which the paper defines as the record of “what has happened in the simulation as a result of the agent’s attempted action.” The GM also maintains the grounded variables, advances the clock, runs the episode loop, and sends each agent the observations it is entitled to.

The paper’s own running example is the attempt “Alex makes breakfast.” A retail version runs the same way. An agent submits “I buy the TV at 199.” The GM checks the inventory variable, rules the attempt a failure because stock sits at zero, writes that event statement, then tells only the agents present in the store what they saw.

The detail that earns this essay is what the paper says the GM may contain. In a physically grounded world it checks the plausibility of each attempt. In digital worlds it executes the API calls the agents intend. And in general, quoting directly, “the GM can use any existing modeling technique to simulate the non-linguistic parts of the simulation (e.g. physical, chemical, digital, financial, etc).” Read that sentence with 2026 eyes. It is the deterministic scaffold and the semantic layer, specified by DeepMind in December 2023, two years before the synthetic-data market consolidated around the same split. The language model plays the people. Anything with rules runs on whatever machinery already computes those rules.

An agent in this architecture cannot decide that its trade filled. It can only request. That is the entire point.

Two implementation details deserve an engineer’s attention before the enthusiasm. First, the GM is itself an agent, assembled from the same component system of LLM calls and associative memory as the players. Giving rules an address does not automatically make every ruling deterministic. A GM that resolves an open-ended attempt by asking the model for a plausible outcome has centralised the prior, one place to audit and override it, which is real progress of a bounded kind. The deterministic share of the rulings lives in the grounded variables and in whatever classical machinery you mount inside the referee. Second, the GM component API carries a method the paper names partial_state(player_name), the state of the world as one player is allowed to see it. The referee doubles as the information boundary. Each agent receives only its projection of the universe. The full state stays with the referee, which is the property any synthetic corpus that must not leak its own ground truth will come to depend on.

The rule inside the agent

The 2007 paper answers the same question from the other end. Tao Zhang and David Zhang built an artificial market of heterogeneous consumer agents to study the decoy effect, the marketing regularity where adding a third option that nobody chooses shifts the choice between the original two. Their agents carry an explicit core motivation function, an equation combining psychological traits with two kinds of market interaction. The rule sits inside every agent, in closed form, where a referee would be redundant.

What makes the paper worth citing 19 years on is what they got out of that transparency. The decoy effect emerged unscripted, out of thousands of legible micro-decisions interacting, which is the same trick Smallville pulled in 2023 when information diffusion fell out of agent conversations no one choreographed. Emergence is the shared ambition of both eras. The difference is that when the 2007 model produced a surprising aggregate, the authors could open any agent and trace the arithmetic that produced it.

Opposite addresses, same property

Put the two designs side by side and the real axis becomes visible. Legibility, with placement free to vary.

The third row is the architecture you get by default when nobody makes the choice. Prompt a language model to simulate a market, a store, a negotiation, with no Game Master and no explicit decision function. Adjudication still happens. Every time the simulated customer accepts a price, the model’s priors ruled on that acceptance, according to patterns no reviewer can enumerate. Smallville’s agents walked in on each other in a one-person bathroom because the base model’s prior about bathrooms outvoted the world’s actual rule. The same paper’s Isabella accepted party suggestions against her own stated interests because instruction tuning made agreeableness the default ruling. Those are adjudications. Nobody wrote them. Nobody can read them. Nobody can point to the line that made them.

This is why I read Concordia’s Game Master as the more consequential half of the pair for anyone building now. The GM is a place. Constraints, physics, budgets, and market rules have an address, which means a model-risk reviewer has something to review. MAP’s production data shows deployed teams already vote this way with their architectures, fixed sequences, scoped action spaces, wrapper APIs, human sign-off at the steps that matter. Production agents live inside referees. The teams just never used the word.

Concordia’s authors are also more careful about validity than the field that adopted their library. The paper imports the hierarchy of evidence from evidence-based medicine, ecological validity at the top, controlled experiments below it, observational data lower, consistency with prior theory at the bottom. And it states plainly that the evidence bar should scale with what the model will be used for, lower for deciding where to pilot, higher for decisions with real consequences. A simulation guiding real capital sits at the top of that ladder, whatever its demo looks like.

The bottom line

For a builder, the checklist falls out directly. Decide where every rule lives before choosing a model. Anything with an invariant, prices, inventory, physics, permissions, belongs in the adjudicator, whether that is a Game Master component or an explicit function inside the agent. Both work. The 2007 paper and the 2023 library prove the placement is negotiable. The address is the part with no give, because a rule without one is being enforced by pretraining priors right now, in whatever direction the training distribution happened to lean.

For a reviewer, the question compresses to one line. Point to the adjudicator. A team that answers with a component, a function, or a rules engine has a simulation you can validate up Concordia’s evidence ladder. A team that answers with the name of a model has handed the referee’s whistle to a distribution nobody has read.

Concordia puts the rule outside the agent. Zhang and Zhang put it inside. Both give every ruling an address a reviewer can visit. That property, legibility of the adjudicator, is the dividing line between a simulation you can validate and a demo you can only watch.