Each horse gets a score from interpretable features, then scores are converted into probabilities that sum to 100% across the active field (a multinomial-logit / softmax).

• Only starts dated before the race are used — no future results leak in.
• Scratched horses are removed and the field is re-ranked.
• If any horse has fewer than 3 prior starts, probabilities are marked unavailable.
• Validation uses time-based splits and reports accuracy, log loss and Brier score.

The model is not advertised as profitable and does not guarantee winners.