Counterfactual Value Calibration for Poker AI

Offline strength is not enough when a system must behave consistently in live decision loops.

Counterfactual value calibration helps align the value model with the action environment it will actually face during inference.

Why the problem appears

Many systems inherit strong offline estimates but lose stability when:

• the live state is noisier,
• the action space is compressed,
• or the timing budget forces approximation.

What calibration improves

Calibration reduces the gap between:

• what the value model predicts,
• and what the deployment context can safely execute.

That improves:

• decision consistency,
• action ranking quality,
• and the usefulness of downstream control logic.