AutoTrust-JEV-27B v1.1, a 27-billion-parameter text decision model, has reached Hugging Face’s trending list. It accepts a question with a finite set of answers, emits one token restricted to the option codes, and converts the token log-probabilities into a distribution over those choices. The model card reports a 67.31 score on Decision Index 0.3, with its strongest results in tool selection, intent routing, and automation.

One token, finite choices

Classification with a general-purpose language model usually involves generating text and parsing the response into an expected label or schema. AutoTrust-JEV uses a shipped decision engine to format each question, call a vLLM backend, constrain the next token to valid option codes, and convert the returned log-probabilities into option-level scores. It supports multiple-choice and yes-or-no questions with text-only inputs.

Restricting the output removes malformed JSON, unexpected prose, and labels outside the supplied option set. The resulting probabilities can support ranking, thresholding, and routing, although they still require calibration on representative application data. Production systems should set thresholds on held-out data and provide an abstention or escalation path for low-confidence cases.

The model card reports a public Decision Index of 67.31 across 37 indexed benchmarks and 140,178 completed, scoreable requests. Tools and Automation is the highest-scoring of the suite’s five areas at 85.19 skill points.