Pick a preset, then press Run. Click the grid to plant a 5x5 patch of reciprocators (a "peace circle").
Each of the 40x30 residents has one of four conflict styles and plays 10 rounds against each of its 4 neighbours every generation. In a round each side either cooperates (C) or defects (D). Payoffs are those of the repeated prisoner's dilemma: both C = R = 3, both D = P = 1, a lone defector gets T (the "temptation" slider), the exploited side gets S = 0.
g (the mediator's forgiveness).g/2 per round.Noise ε: every intended move is flipped with probability ε (misunderstanding, a message read the wrong way). Two reciprocators then fall into echoing retaliation; forgiveness breaks the echo. For two reciprocators the long-run state is a 4-state Markov chain on (my last move, your last move) with P(C | other's last = C) = 1-ε and P(C | other's last = D) = g(1-ε)+(1-g)ε; its stationary distribution gives the exact mean payoff per round, which the check panel compares with simulated play.
After each generation every resident looks at one random neighbour and copies that neighbour's style with the Fermi probability 1/(1+exp(-(πj-πi)/κ)), κ = 0.1; a small mutation rate adds new styles.
Lesson: forgiveness helps up to a point and then backfires - when almost every defection is forgiven, pushing pays. Exact numbers depend on the illustrative payoffs.
Honest simplifications: payoffs are illustrative numbers, not measured; real conflicts are not a one-shot matrix, and people hold more than four strategies. The model shows a mechanism (reciprocity + forgiveness protects cooperation under noise), not a prediction for any real dispute.