Pick a scenario and click "Run decisions"…
One forward pass per request: every field is read from its own answer slot, no tokens are generated. prob is the model's probability for the chosen value after temperature scaling. Scenarios adapted from the jev-on-a-laptop study (MIT).