Typed decisions
Give the model context, fill in typed questions, and inspect the decisions with their probabilities. Every released version answers the same question, so you can see where they differ.
Everything the decision depends on. The model reads this and nothing else — paste a ticket, an alert, a diff, or leave the example.
All questions share this context and are answered independently, so asking ten costs one read of the state.
What you want decided about it. Each question is answered on its own, and comes back as a probability rather than a sentence.
noul
one probability that the answer is yeschoice
a probability for every option you listscore
a probability for each level, and the weighted position on the scale
Editing does not re-run the model: a pass costs one read of the state per version, so you spend it when you choose to.
One row per option. Each selected version draws its own bar, so where two releases disagree is visible without reading the numbers.
JSON sent when you press Run. Copy it to call the same endpoint from
your own code. The released server speaks the Jev contract at
POST /v1/systemone; this page adds the version list.
Text examples are held out from LocalLLaMA/typed-decisions, which no version
is trained on. The image examples (a chart, a screenshot, a receipt) are drawn by
serve/demo_images.py, so their answers are known by construction. Teacher is the reference answer, shown only while an example is unedited,
because a document you write has no right answer to check against. Bars are the model's
probability for each option, with rules at 25, 50 and 75 per cent.
The page's layout follows the
Jev AI Community playground, with thanks.
Built on an HP ZGX Nano AI Station with the NVIDIA GB10 Grace Blackwell Superchip, provided by HP and NVIDIA.