Decision Lab

The Decision Lab asks a decision model typed questions about a situation and shows the probability behind every answer.

Sidebar item: Decision Lab (below Cache Lab) Route: /decision-labAdded in: v0.7.0 (October 2026) Requires: Ollama 0.35.1 or newer, and a decision model (nimble, tev1, clef-flash or clef)

What's the point?

A decision model does not write text. You give it a state (a ticket, a message, a log line, a document) and one or more questions, and it returns:

Question type	What comes back
Choice	The chosen candidate, a probability for every candidate, and a confidence
Yes / no probability	A single probability that the condition holds
Score on a scale	The expected position on your ordered scale, a probability per step, and a confidence

The probabilities are the model's own output distribution over the candidates, not a number parsed out of prose. That is what makes them worth looking at: a chat model asked "is this urgent?" says yes; a decision model says 0.99, and for a borderline ticket it says 0.55, and you can see the difference.

Ollama serves these models on /v1/systemone (added in 0.35.1). They report decision as their only capability, so LLMxRay keeps them out of every chat and embedding selector and lists them here instead. Sending one a chat request does not fail cleanly, it answers with fragments of its own schema, which is exactly why the filter exists.

Using the page

1. Pick a decision model. If none is installed, pull one from the Models page; tev1:0.8b is the smallest.
2. Describe the state. Every question is asked about this one text.
3. Edit the questions. Each has an answer key (what the result comes back under), a type, instructions, and for choice and score types a list of candidates:
   • Choice: one per line as name: description.
   • Score: one per line, lowest first. The order is the scale.
   • Ollama accepts 2 to 26 candidates and 1 to 64 questions per request; the page tells you before sending when a question is outside that.
4. Decide. The answers appear on the right as bars, the chosen candidate highlighted, with input and output token counts and the round-trip time.

The page opens with a worked sample: a support ticket asked three ways, one question of each type. It is the request that was run live while verifying Ollama 0.40.3.

Reading the answers

• Choice: bars are sorted by probability. The highlighted one is the model's choice. Confidence is the model's own figure, reported alongside.
• Yes / no: shown as a yes and a no bar that sum to 1.
• Score: bars stay in scale order, because position is the meaning. The highlighted step is the most probable one. The score number is the expected position (probability-weighted), so it can fall between two steps: a 2.37 on a four-step scale means "major, leaning toward full outage".

What this page does not do

It does not stream (the endpoint never does), it does not keep history, and it does not send images yet, although clef and clef-flash accept them. It is an observatory for what a decision model actually returns, nothing more.