Run it yourself
Interactive predictor
Pick any two fighters with at least one UFC fight in the replay and the model runs live in your browser — no API call.
The headline model is an ensemble (logistic + symmetrically-augmented gradient boosters), the best on the scoreboard. This uses the served logistic instead — it's within ~0.003 log-loss of the ensemble and is the only one that can run entirely client-side (a dot product, no model runtime).
Served model not available yet - run python -m etl.ml.train --served logistic_l1.

