Training data consists of live hand histories and session records collected across Las Vegas 2/5 and 5/10 cardrooms over ten years. Key statistical targets — pot size distributions, win/loss pot asymmetry, VPIP/PFR/WTSD/WSD rates, street-by-street pot growth, and showdown frequencies — are extracted from this corpus and used as calibration ground truth. Simulator output is continuously benchmarked against these empirical distributions using KL divergence and KS statistics. Models are retrained when drift is detected.

Pot size prediction is handled by a mixture density network (MDN) rather than a point estimator. The MDN outputs parameters of a four-component Gaussian mixture — capturing the multimodal reality of poker pots: steal pots, single-raised pots, 3-bet pots, and all-in pots. Separate conditional MDNs are trained for winning hands, losing hands, and hands that end without showdown, preserving the win/loss pot asymmetry that is the primary differentiator between winning and losing players. Bluff and cooler scenarios are handled by dedicated specialist models rather than forcing rare high-variance events into the general distribution.

Win probability and showdown likelihood are modeled using LightGBM classifiers over a ~50-feature input space covering hand strength, board texture, player tendency cross-features (bluff_freq vs fold_to_bluff, aggression delta), position, stack depth, and game structure. The LightGBM leaf embeddings serve as input to the MDN heads, coupling tree-based feature learning with distributional output. System thresholds — fold equity cutoffs, pot control triggers, barrel frequencies — are not hand-tuned. They are optimized automatically via Bayesian optimization over a surrogate model of the calibration loss, converging in 50–200 simulator evaluations rather than exhaustive grid search.

The Monte Carlo engine deals real cards, computes per-street equity via sampled board runouts, and resolves actions through player behavior models — producing both aggregate statistics and inspectable hand histories. Aggregate stats are automatically compared against real session data; hand histories are periodically reviewed by human raters using pair wise preference judgments (which of two simulated hands is more realistic). These preferencestrain a lightweight Bradley-Terry reward model that scores simulation realism at scale without requiring hand-by-hand human review. The result is a self-improving loop: real data anchors the statistics, human judgment anchors the narrative realism, and Bayesian optimization continuously closes the gap.