Short answer
For a bounded coding decision, pixel classify asks a model you configure and returns one probability per label. Ollaya reports typed-decisions accuracy of 0.722 for winnow:e4b and 0.738 for Jev. Separately, it reports 89 ms for five local questions on an RTX 4090 and 236–276 ms for a hosted Jev request. These latency setups are not comparable. The rest of Pixel needs no decision model.
Side by side
Published figures, different samples. Pixel's own measurement set beside Jev's published figure: not a head-to-head, and the samples differ.
| Measure | Pixel | Jev |
|---|---|---|
| Typed-decisions accuracy | 0.722 | 0.738 |
| Latency | 89 ms for five questions (RTX 4090) | 236–276 ms, hosted median request |
| Model | winnow:e4b on Ollaya | TypeSafe Jev |
Every figure above is on the benchmarks page, with its sample size, its method and the cases where Pixel loses.
Where Jev wins
Its 0.738 typed-decisions accuracy is slightly higher than winnow:e4b's 0.722.
How they differ
Jev is a model built for one job: settling a coding question with a bounded set of answers. pixel classify puts the same kind of question to a model you configure and returns one probability per label. It is an optional add-on: nothing else in Pixel calls a model.
Choosing
Pick Jev when its slightly higher typed-decisions accuracy matters more than local latency. Pick pixel classify to keep the choice of model, the bill and the data path in your hands, next to an index your agent already queries.