The AI players are a neural network trained by self-play: one shared policy plays all four seats
for millions of hands and improves with reinforcement learning (PPO), approximating an equilibrium
strategy. Separately trained "best response" agents could not find a way to exploit it.
EVs are computed from your information only. Possible hidden hands are sampled consistently with
everything you've seen and weighted by how likely the AI players were to bid and play as they did with
those cards; each option is then played out by the policy. From trick 4 on, every layout is enumerated
and the EVs are exact. An EV is the expected point differential for your team on this hand
(makers 1, march 2, alone march 4, euchre 2).