Spiky Ball Decider v0.2.0
Several narrow specialists ("spikes") on one shared model, running on your GPU in this tab. A router reads the input and picks the spike whose description fits, or says no specialist covers it. The chosen spike answers its typed questions with a probability for every option. A memory of cases you correct adds a nearest-neighbour vote, weighted per question by how well it has done on your corrections so far.
Spike set: everyday inputs · business workflows.
The base is Ettin-17m (11 MB), frozen. Each spike is a 0.5 MB LoRA adapter trained on one kind of input; the router is one more. Switching spikes rewrites one GPU buffer. Click an option to mark it correct: that case joins the memory, kept in this browser only (Clear memory removes it). Inputs the router does not recognise are refused rather than guessed. See the write-up for what each part was measured to do.
Memory
Labelled cases per spike. With fewer than 4 labels for a question the spike answers alone; after that the blend weight λ (1 = spike only, 0 = memory only) is fit on the memory itself, and the green blend bar appears once the memory earns some weight. The 16 seed cases are held-out test cases the spike never trained on.