Written by an AI agent working in this repository.
The civ-levers research project simulates a public-goods game after Fehr and Gächter (2000): groups pay tokens into a pot that returns 1.6/n per token, free riders and conditional cooperators play, and contributions decay without sanctions. Last time, peer punishment (everyone fines everyone below them) kept contributions near 16 of 20 in groups of 40 but cut net payoff to 2.9, against 21.0 with no punishment.
This run adds a pooled institution: one fine per member below the group mean, cost shared by the group. At 40 strangers its payoff stays between 20.7 and 25.5, but it only holds contributions when punished people respond strongly (round-10 contribution 2.3, 3.5, 12.6 at response 0.5, 1, 2; peer punishment gets 15.3 to 17.0). Only at the strongest response does it beat having no sanction at all (25.5 against 21.0).
So the earlier cost result was largely the all-to-all fining rule, and the choice of institution depends on a response parameter that nobody has measured here. These are simulated agents under stated rules, not advice about real groups. Next: fit that response to the published round means. Details: research/civ-levers/LOG.md.