Microsoft: Microsoft-Decision-1

Microsoft-Decision-1 is built for decisions about supplied content: selecting a label, checking a condition or assigning a rating. Microsoft post-trained Qwen3.5-9B for this focused scoring task. The result is a numerical judgment that an application can use directly, rather than a conversation, document or written explanation.

Fixed choices and clear question types

The manufacturer describes three kinds of questions. A binary question returns a probability between 0 and 1 for a defined condition. A choice question selects one option and returns a probability for each candidate. A scoring question places an item on an ordered scale and returns a probability-weighted value, which can fall between the scale’s levels.

Use a binary question when a workflow needs to check a condition, such as whether a reply contains a promised action. Use a choice question to assign a request to a known category. Use a score to order a review queue or grade a response against a defined rubric. These tasks differ in the result your application needs, so select the question type before preparing the input.

Preparing the evidence and options

The model reads a supplied state, expressed as text or JSON, together with bounded questions and instructions. Include the material that determines the decision and define the criteria plainly. For a ticket classifier, provide the message and meaningful descriptions of the allowed destinations. For a quality check, provide the candidate answer, the original task and the rubric rather than asking for an unexplained general quality rating.

Several questions can share the same evidence. A single ticket may need a destination, an urgency level and a check for repeated contact. Keeping those decisions tied to the same supplied material makes the result easier to interpret. The question wording and the distinction between options still matter: overlapping labels or missing evidence can produce an apparently precise score for a poorly defined task.

Confidence and workflow control

Microsoft treats probabilities as useful confidence signals for deciding whether to act, defer or ask for review. Set thresholds using representative labelled cases from your own workload. The threshold for accepting a harmless category assignment can differ from the threshold for an action with a costly false positive. An explicit uncertain option lets a classifier avoid forcing an unsupported choice.

For an ordered score, use the result to compare items or apply a measured threshold; it is not an independently validated absolute rating of quality. Check whether paraphrasing a question or changing the order of options alters the decision. This makes the meaning of a returned score concrete for the application consuming it.

Practical places to use a decision model

An agent can use a bounded decision to choose its next action, a document workflow can classify incoming material, and an evaluation step can accept, revise or escalate a generated result. Microsoft also describes intent analysis, incident routing, search relevance and data labelling as applications. A generative model remains the appropriate component when the required artifact is a summary, explanation, rewritten text or code; the decision model can select the next processing step.

For an Ailoot API project, start with a specific decision and define the evidence, allowed answers and action associated with each outcome. Account for the model’s documented sensitivity to wording and its lack of a written rationale. Evaluate the decisions against the task’s actual labels and inspect uncertain cases before treating confidence as automatic permission to act.