The bottle that became an airplane

The live vision pipeline was running smoothly. The bottle, apparently, had other ambitions. Held vertically, it was repeatedly classified as a surfboard. Turned horizontally, the same object was repeatedly classified as an airplane.

The horizontal test recorded 287 airplane detections during a 291-frame run, with no dropped frames. Those are detection counts, not 287 separate objects. Some predictions were confident, and one bounding box covered nearly the full image.

We did not recover a proven explanation of the model’s internal reasoning. Shape, scale, and orientation were plausible contributors. What the test did establish was simpler: smooth acquisition and repeated confident predictions did not guarantee the right label.

That failure supported a central design decision for Ash. Raw detections would pass through a validation layer before becoming events that could influence behavior. The bottle test remained a classification failure and a useful diagnostic result.