Examples
Each one was run before we put it here; the card shows the recorded answer. Click to load it and run it live, in your browser.
Long cases the model never trained on
One situation, several questions about it. Decima answers each question live in your browser; next to it is the answer of the 26B teacher that labelled the training data — including where they disagree.
Reorder the options. The answer stays put.
Each option is scored against the situation on its own, so its position never enters the computation. Shuffle as often as you like; the numbers below count every try in this session.
Automate the sure cases, hand over the rest
Because the probabilities are calibrated, the threshold is a dependable dial: raise it and fewer tickets are automated, and fewer mistakes get through. Twenty realistic tickets in eight languages, expected team written down before the model saw them. Your browser scores them once; after that the slider is instant.
Twenty tickets we wrote — a demonstration of the mechanism, not an automation rate. Pick your threshold on your own labelled tickets. A threshold is not an off-topic filter: a message about something else entirely can still score high on one of your options.
Measured on your device, now
Time one decision in this browser tab with 4, 20, 77 and 150 options (WebAssembly, option encodings cached). Cost grows with the number of options; the dotted line is our x86 laptop reference, native ONNX Runtime on one core.
Reference, x86 laptop (Intel Core Ultra 7 155H, one P-core, int8, native ONNX Runtime): 20 ms with 4 options, 42 ms with 20, 83 ms with 77 (model-card option sets). With short labels, as plotted: 16 ms at 4 options and about 1 ms per extra option; 1,000 options take 1.06 s. A browser tab runs the same int8 model through WebAssembly, usually on one thread, so expect it to be slower.