Swapping the backing model moved every measured outcome in a 100-agent town economy

Conservation here comes from the loop shape. The published pulse resolves every agent's location, calls each agent's model in turn for a decision, hands the whole batch to one settlement function, then asserts that the settled amounts sum to zero before any agent state is updated ^[15]. No agent writes a balance while another is still deciding. That ordering, not a lock, is what satisfies the stated requirement of balancing every transaction across 100 concurrent agents without a central ledger lock ^[4]. The assertion sits after settlement, so it detects an unbalanced batch and cannot prevent one.

Two validation layers sit around that. A live validator checks monetary conservation at each pulse, and an offline pass recomputes each agent's wealth from its signed, append-only transaction history ^[13][12]. The second layer exists because the pulse assertion only sees the batch in front of it. The team says every headline number was verified twice and has released the full run corpus for reanalysis ^[13][14].

The cash-transfer arm is the one I would read first, because it drives the same question from the opposite direction. NPR 5,000 went to 20 of the 100 agents; 311 pulses later, 96.7 percent of it was still held, and marginal propensity to consume came in at 3 to 4 percent on two independent measures ^[10]. Those two results agree: 3.3 percent of the transferred cash left the recipients' hands, inside the reported band ^[5]. On the supply side, 0.3 percent of 3,981 menu items works out to about a dozen items ever repriced ^[2].

For any of this to describe a different system, the tool surface and the one-decision-per-pulse cadence have to look similar, and the model has to be in the same neighbourhood. Both families tested, described in the write-up as GPT-4 class and Claude class, showed the same price rigidity and wage stickiness ^[19]. The write-up calls this the first published multi-agent economic simulation to run past one or two weeks into 26-week territory ^[3], which I believe mostly because week 26 is expensive.

Deadlock detection is where the account thins. The metrics list includes price change frequency, wealth concentration by Gini and Spearman rho, tool call success rates and monetary conservation violations, and the write-up says these surfaced the economic deadlock before the simulation diverged ^[20]. There is no threshold given, no trigger, and no description of what the harness did once it fired. Handling agent-set prices that prevent market clearing is listed as a design requirement ^[6]; the detector behind it is not shown.