Load the models you and others trained, then play them against each other. Every model was trained under the same house rules, so the fight is fair; the file records each one's weight class and training, and a warning shows if you mix classes.
The goal: beat the simple lookup bot (basic strategy, flat bet) — the naive player. Download it and drop it in as your opponent: SimpleLookupBot. For tougher benchmarks, try the counters BlackJackKing and BlackJackSage. Why beat the bot and not the house? See the Findings.
Play many sessions and rank by results. Duplicate deals every model the identical cards in isolation, the fairest measure. Shared table seats them together over one shoe, like a real casino, where one model's hit changes the next model's card.
One shared shoe, dealt hand by hand. Turn on "skip the obvious hands" to fast-forward the routine ones and slow down when a model makes a big bet, splits, doubles, or a blackjack lands.