Approach

Each lane asks a model one question per frame and gets one probability per action.

Training data

orange = trained · gray = frozen

No training. Read each option's score. Our floor.

Kev-style. Train a small add-on and a head.

Closest action vector wins.

Reads pixels. Sees things, not where they are.