My own experiments from the Apart Research collusion sprint: a published collusion detector run on 295 games between Llama and Mistral agents in Rock-Paper-Scissors and a double auction.
It finds seeded pairs 16 of 16, false-alarms when two copies of one model play among strangers, needs thresholds calibrated on honest games, and cannot see three colluders. The full note has three interactive figures.
The full note has the complete argument and its interactive figures.
Read the full note