Rust game engine
The engine is written in Rust and exposed to Python through PyO3/maturin. Money is an i64 in milli-big-blinds, so no floating-point chips exist anywhere, and every hand is checked against chip-conservation invariants. Hands are deterministic and replayable. The hand evaluator is a sort-and-categorize design packed into a u32 rank instead of a 130 MB lookup table: no static data, fully auditable, and still within the microsecond budget, with optimization kept behind a stable evaluate_5/evaluate_7 API.
Seat-attention network
A network of about 1.4M parameters attends across seats, so the controlled seats are represented together rather than as separate inputs. It is the policy and value core the blueprint and the search both rely on.
Blueprint, then exploitation
Training starts with a DREAM/CFR-family blueprint, which gives a provable floor on exploitability. PPO then fine-tunes against weaker opponents with a KL anchor back to the blueprint. One method alone gives either safety or exploitation; the combination gives both.
Joint real-time search
At decision time, CFR+ search refines the blueprint. A coalition range gadget computes the team’s counterfactual values jointly. Summing per-seat values would be wrong, because card-removal and blocker effects cross between the coalition’s known cards; the spec treats a per-seat sum as a build-breaking defect.
Opponent modelling
Beta-binomial statistics track each opponent’s tendencies, so exploitation adapts to the players actually at the table instead of a fixed population.
Built like flight software
The system is built spec-first against about 155 ordered, verifiable tasks. Each task has a contract, the engineer and auditor roles are kept strictly separate, and phases close only on a human-signed git tag. A correctness-critical AI system is built auditably rather than in one pass.
