An evolutionary tournament with noise and priced verification — can verifiers survive, and at what cost?
You played The Evolution of Trust. It ended in fog: a little miscommunication, and cooperation ate itself. This is the sequel. The world gets one new ability, paying to see the truth, and one new question: when is the price worth paying?
Cooperation between rivals often fails not because anyone is evil, but because nobody can tell what anyone actually did. Verification is the technology that fixes that. It is never free.
Stage 1. The world you know
A one-screen recap, then we move on. 25 players. Every pair plays a 10-round match. Each round, both secretly choose: cooperate or defect. Both cooperate: 3 points each. Both defect: 1 each. If one defects on a cooperator, the defector takes 5 and the cooperator gets 0. After everyone has played everyone, the 5 lowest scorers are eliminated and the 5 highest are copied. That is one generation. We run 30.
Information is perfect here. What everyone sees is exactly what everyone did.
Predict first: Five characters, five of each. Who owns this world by generation 30?
The run button unlocks once you commit.
Under the hood: rules, numbers, and honesty
The payoff matrix
Each round both players secretly choose, then reveal. Points per round:
| They cooperate | They defect | |
|---|---|---|
| You cooperate | you 3, they 3 | you 0, they 5 |
| You defect | you 5, they 0 | you 1, they 1 |
The noise model: misperception, not slipped hands
Every round, each player separately records what they think the other side did. With probability equal to the noise setting, that record flips, independently for each observer each round. The true action stands: payoffs are always computed from what really happened. Strategies decide using only their recorded history. So noise here never changes anyone's behavior directly. It changes what they believe happened, and their rules react to the belief. This is why verification can help at all: an act that really happened cannot be undone, but a wrong record can be corrected.
The verify mechanic
A verifier's records never flip (at reliability 1.0). It perceives true actions. In exchange it pays the verification cost every round of every match it plays, regardless of outcome. Verification here is always-on, accurate, and priced per round. The free-play extension adds a reliability setting r: the verifier's record is correct with probability r, so r below 1.0 makes verification imperfect.
The replication rule
25 players. Each generation, every pair plays one 10-round match (free play can change the length) and scores are summed. The 5 lowest totals are eliminated and the 5 highest are copied. Ties are broken by a seeded random draw. The elimination count stays 5 even if you shrink the population in free play, so tiny populations evolve violently.
Calibration: what we checked, and what we changed
Before shipping, the engine was run headlessly for 20 seeds and 30 generations per stage. Pass rates against each stage's expected outcome: Stage 1, reciprocators hold a majority by generation 30 in 20 of 20 seeds. Stage 2, defectors reach at least half the population in 19 of 20. Stage 3, verifying reciprocators are the largest family in 19 of 20. Nothing on screen is scripted: the stages replay whatever the engine produces.
Two parameters differ from the original design brief, and we changed them in the open. The brief asked for 10 percent noise in Stages 2 and 3. At 10 percent, with 10-round matches, misperception turned out to be too mild to break reciprocity (a single misread between two Copycats costs an alternating echo, not a collapse), and the measured value of true perception, about 0.13 points per round, was below the 0.2 price, so verifiers always lost money. Stages 2 and 3 therefore run at 20 percent noise, and Stage 3's cast swaps one Copycat for one V-Copycat. Payoffs were never touched and no seed is special-cased.
The default seed is 3. Of seeds 1 through 20, 18 pass all three stage checks; seed 3 was picked among them because its runs are representative and easy to read (Stage 3 shows the plain reciprocators rallying before the verifiers overtake them). Any seed gives an honest run; type your own below.
Known sensitivities
The corrosion threshold for noise sits between 15 and 20 percent: below roughly 15 percent, reciprocators usually survive the fog unaided. Outcomes are winner-take-all (one family usually sweeps), so near a boundary, such as verification cost 0.22 to 0.25, the same settings can succeed on one seed and fail on another. Longer matches favor verifiers but weaken the Stage 1 story. V-Cooperator never outperforms Cooperator at any positive cost: it pays for information its rule never uses.
Every number on screen comes from these rules.
Inspired by Nicky Case's The Evolution of Trust (ncase.me/trust). Built for XLab Tracks · Verification, Existential Risk Laboratory, University of Chicago.