Motion Eval
How the Motion Eval Works
This is a live deployment eval that runs on you. We stream your webcam, take the per-region pixel difference between frames, and call wherever you moved a "hit." No model, no inference budget — just you, physically, in the latent space of your living room.
- Allow the camera (or skip it — pointer/touch is a perfectly valid fallback)
- Clean examples (cyan) drift in: move into them to catch the data
- Adversarial examples (pink) also drift in: do not touch them, or they poison your run
- Chain catches for combo multipliers; lose all your integrity and the eval halts
Why Am I Sweating?
Because the difficulty ramps. Token throughput climbs, examples get faster and denser, and your reflexes were never RLHF'd for this. The webcam feed is mirrored and faded so you can see yourself flail in real time — fully on-device, nothing leaves the browser.
Slop Fact: Frame-differencing motion detection predates deep learning by decades, which makes it the only part of this stack that actually generalizes. Your webcam stream never leaves the page — we couldn't afford the GPU to do anything useful with it anyway.