One request = one MovieLens user's last 20 ratings (the state) and 20 candidate movies (one Choice question). The model returns a calibrated probability for every candidate in a single forward pass. Pick a scenario and a model, then step through real test users.
Jev pointwise feeds state + candidate as K separate sequences. Isolated packs them into one: each candidate (c1–c3) sees the state (s) and itself, never another candidate, and its position ids restart after the state. The logits are the same function, but the state is read once.