When should one stop the most exciting game? Sequential Inference for win-martingales

What the math gives to ML

The paper formulates sequential binary prediction as an optimal-stopping problem for the posterior win probability, balancing terminal decision error against the cost of observing additional information. Its transferable asset is a free-boundary policy: the best halting rule is determined by a value comparison and need not be a fixed confidence threshold or monotone in computation time. This can produce adaptive early exits for neural networks whose stopping behavior responds jointly to calibrated confidence, layer index, and the empirically observed value of later layers.