Hunch.
Judgment, in one pass.

From a situation to a signal

Many useful decisions have a known set of possible answers. A support message needs a route. A claim needs to be checked against evidence. A piece of text needs to be scored against a rubric.

Hunch takes a state, a question, and described candidates. It returns a probability distribution over those candidates in a single forward pass. There is no generated explanation or token-by-token answer to parse.

The application keeps control: it can select an answer, apply a confidence threshold, combine several judgments, or hand the case to a person.

Three ways to ask

Choice

A probability for each supplied candidate. The interface supports 2–255 candidates.

Boolean

The probability of a proposition being true. The caller chooses the decision threshold.

Score

A distribution over 2–10 described, ordered levels, with an expected level available through deterministic post-processing.

These are interface capabilities, not a claim that every task or candidate count has the same measured quality.

What the preview explores

The current 1.7B research preview adapts a pretrained backbone for routing and triage over text or JSON states. Its intended applications include support intent classification, evidence verification, paraphrase detection, and selected text attributes.

The research asks how well those judgments transfer to new tasks, how candidate descriptions affect performance, and whether the probabilities remain useful outside the training families.

Uncertainty is part of the result

A normalized distribution is not automatically a calibrated one. A model that performs well on familiar examples may be overconfident on a new task.

The current model card documents out-of-family calibration and generalization limitations. The preview is not production-validated. Evaluate it on your own data, choose thresholds for the actual workflow, and keep human review where the consequences require it.

Release evaluation is ongoing. This note intentionally makes no latency, superiority, or production-readiness claim.

Research in the open

The project is working toward releases that pair open weights with reproducible data recipes, evaluation tools, and a model card. Each release should make its training scope, evaluation conditions, and known failures inspectable.

Download links will accompany an actual released checkpoint. This page introduces the research direction and interface; it does not announce general availability.