Hold up a real cup and a photo of a cup. Both look like “cup” to the model. Toggle Afferens: off, the arm grabs whatever flickers highest — often the picture. On, it verifies (stable across frames) and grabs the real one.
“Verified” here is a real measurement (frame-stability), not a confidence threshold. Depth unavailable — running on temporal stability (works on any phone).
Agent log
// the agent’s real reasoning streams here…
The honesty: the OFF side is the model’s genuine raw output — it really does pick the photo, because the photo really does score high. The ON side adds real checks (frame stability). Fool the checks and it’ll honestly still fail — verification is a real signal, not a magic trick. That refusal is the product.