New "System One" decision model from the author of Laya: VegaML. Into the benchmark it went, on my M4 16 GB MacBook.How ...

New "System One" decision model from the author of Laya: VegaML. Into the benchmark it went, on my M4 16 GB MacBook.How it works: a frozen Qwen3.5-0.8B reads the question, then a 14M-parameter "physics engine" lets a ball roll into the valley of the right answer. No text is generated, only probabilities, plus abstain flags and conformal sets.My 80 routing prompts:- Zero-shot: 51% topic accuracy (Jev: 89%), confident on 1 prompt out of 80, 1.3 s per request on the M4 (4 questions, so 4 passes)- With fit() on 40 labelled prompts, about a minute and no training run: 65% topic, and risk beats Jev (64% vs 61%)- The abstain/unbound/conformal signals didn't tell the router anything its confidence threshold doesn't already knowNot ready to route my prompts, but it's the only model I've tested that learns something useful from a handful of labels in a minute.Also: Laya 0.4.2. Eight releases, same weights, identical answers on all 80 prompts. And a DeepSeek price change on OpenRouter moved up to...