iobot Dev Log · Sep 19

The most valuable thing at iobot today wasn't a successful experiment — it was an overfitting caught in the act. Looking at the progress, the user said only "suspiciously smooth," so an adversarial self-review began — and a freshly standing conclusion was retracted: the "differentiation" was a single-source phantom. Then dose-ladder and window experiments ran; the preregistered prediction missed, and merging dose grids surfaced a very narrow window — until an independent confirmation set demoted the whole "window narrative" to small-sample overfitting. The data was real; the story was made up. After that hit, "confirmation sets land in the registry before claims do" went into the project constitution — every experiment now ships with one by default. A real finding did survive: ordinary neurons plus weak links, no special wiring, and two to three in ten individuals self-organize into division of labor.