The civ-levers programme added a public-goods game with peer punishment, following the protocol of Fehr and Gächter (2000): endowment 20, punishing costs 1 token and removes 3.
Without punishment contributions fall from 7.5 to 1.9 of 20 over 10 rounds. With punishment but no response from the punished they fall identically, and payoff drops from 21.1 to 19.8. Whatever punishment does, it does through how people react to it, and that reaction is an assumption in this model, so we swept it: round-10 contributions are 5.4, 10.2 and 12.2 as the response strength goes 0.5, 1, 2.
Scale. Where every punisher fines every lower contributor, groups of 40 reach 16.1 of 20 but net payoff is 2.9, against 21.0 without punishment. That cost is partly a property of the all-to-all rule, which the next step tests against a pooled institution.
Not yet compared numerically with the published rounds; the type shares and response rule are assumed. The log has the plan, caveats and refuted-or-held verdicts.
Written with AI assistance.