Why is Split EV an approximation?

Split EV is computed by processing CDZ (Card Distribution Zone) data through an ML model, rather than exact analytical calculation. As a result, the values shown are approximations.

Due to f64 floating-point precision limits, the displayed Split EV may differ slightly from the true theoretical value. This is a known limitation. The error is not large enough to affect optimal decision-making, but perfect accuracy cannot be guaranteed.

Split is the most expensive action to evaluate exactly. After a split each new hand can draw, double, and in many rule sets re-split, branching into a large tree of outcomes that must each be weighed against the live shoe. Doing this analytically for every possible draw sequence is far more work than evaluating Hit, Stand, or Double.

To keep the app responsive, Split EV is produced by a machine-learning model trained on exact CDZ data instead of the full analytical recursion. The trade-off is speed for a tiny accuracy cost — on the order of 0.08% RTP error — while running roughly 1000× faster than the exact computation. For decision-making this error is negligible.

In practice you can trust the Split EV figure to rank actions correctly; the approximation almost never changes which play is best. Treat the last decimal place as indicative rather than exact, and if you need the precise analytical value for research, remember the displayed number can differ slightly due to both the model approximation and f64 floating-point limits.