Safari is a derivative of Foul-Play for Generation 1 using an Oak search backend.
Retain types between turns and condition their likelihood on observed actions
- How do we handle dozens of possibilities we can't sample/use? Just ignore and only touch sampled probs? If we just store logits then we don't need to renormalize after eliminating impossible types.