Interaction study
Ten drives, two ways to say yes.
The bulk review pattern behind my MileIQ work, instrumented. The same ten suggested classifications under two confirmation models. The only variable is what the default does. Run both and the counters keep score.
Pick a model · tap chips to confirm or flip · marks the model’s suggestion · the footer is live
The numbers show that the default does most of the work. Flipping exceptions inherits the model’s accuracy as its error floor, so it only earns its speed when exceptions are cheap to spot, which is why the shipped feature paired bulk accept with a visible recommendation mark and an undo that takes one tap. Effort moves out. Vigilance has to move in.
SERIES · PROTO_02: Set the line → · PROTO_03: One field → · PROTO_04: Swipe to classify →