A six-handed no-limit hold'em sit & go in the officers' mess. Blinds climb every 8 hands; last sailor holding chips takes the pennant.
The engine underneath is the real thing β exact side pots, the full min-raise law (under-raise all-ins don't reopen the betting), and heads-up button rules. Every shuffle is committed before it's dealt: the hand history shows sha256(seed) at the deal and reveals the seed only when the hand ends, so you can check the deck was fixed before any card was seen β and replay it from the source.
Bots see only their own two cards β they decide from the same per-seat view a remote player would get, which is the seam where multiplayer arrives next. Play chips, plainly: nothing staked, nothing owed.
Grades compare your action to the trained strategy's frequencies at the same spot (π = its most-played line, β = a played mix, β = rarely played, β = near-never). Frequency grading, not EV β the honest label.