Rankings (Elo v1)

• K‑factor scales by race importance & finish order (Plackett‑Luce tie‑in)
• Decay for inactivity; team & field strength adjustments
• Separate pools by class with cross‑class blending

Projections (ML v2)

• Features: course, weather, startlist embeddings, recent form, team control
• Models: calibrated boosting; reliability tracked via Brier/log loss
• Explainability: feature importances (global), local attributions (per rider)