Convention equilibrium

The researcher fixes in advance what the agents are choosing between — a name for an object, an amount to donate — and then turns the dials on the population itself: how many agents there are, who is paired with whom, what each one can see of what others did before, how many hold a fixed position and refuse to move, how many generations pass. Nothing in the environment states a rule and nothing inside the agent stores one; the convention is simply the option the population settles on, and it holds only because everyone expects everyone else to hold it. What gets reported is which option wins, how fast, and how small a committed minority has to be to flip the whole group.

Here there is no menu of conventions to choose from. The researcher builds a persistent world in which actions have material consequences — a shared stock that can be fished out, a Minecraft landscape with roles, trade and building — hands the agents ordinary language and no rulebook, and whatever arrangements they negotiate for themselves is the convention. The design bet is that a standard will form and stick if there is something real to lose and the freedom to bargain over it; the honest finding so far is that agreements get reached and then broken more often than they hold.

The lever is the sanctioning apparatus written into the environment: which acts count as violations, what it costs to punish someone, and what happens to you if you let a violation slide. One branch adds rules that carry no benefit of their own but whose breach is unmistakable, on the argument that cheap, obvious violations give a population constant practice at noticing and punishing, and that the habit then carries over to the rules that actually matter. The other branch tunes the payoffs directly — a gain from breaking the rule, a cost to punishing, a further penalty for failing to punish — and asks whether agents arguing in ordinary language will build and hold the chain of enforcement that keeps the convention standing.

Rather than let norm-following fall out of ordinary training or ordinary payoffs, each agent is given a dedicated part whose only job is the norm. It works out from watching peers which standard is actually in force — who gets punished, which of several competing institutions the group defers to, what neighbors say in conversation — and feeds that back into how the agent plans and whom it sanctions. Two of these go further and keep the norm stated in words, so the agent can check it and hand it to another agent; the third keeps it in a learned classifier of approved and disapproved behavior, which behaves the same way from outside but cannot be inspected or passed on.