Liquid AI Introduces Decision Models for Structured Decisions

Summary

Liquid AI’s documentation introduces Decision Models, a model class designed for structured decisions rather than token-by-token text generation. A call evaluates a state and returns calibrated probabilities across fixed outcomes with zero generated tokens. The API exposes three primitives: Noul for a yes-or-no probability, Choice for a probability distribution over unordered options, and Score for a probability-weighted position on an ordered rubric. The documentation advises using Noul for binary gates, Choice when code branches on categories, and Score when it compares a value along a continuum; a Noul probability near 0.5 expresses uncertainty, while Score represents degree or severity. Developers create a Liquid API key, call the `systemone` endpoint, select a model such as `d1:free`, and provide text or JSON state plus one or more named questions. Each question includes a type and instructions, while Choice and Score questions also define their options or levels. Multiple question types can be evaluated in one round trip, allowing an application to classify, route, and prioritize the same input. Responses include one answer per question: Noul returns a probability, Choice returns a selected option, full distribution, and confidence, and Score returns a numeric value, distribution, confidence, and rubric legend. Examples show complaint and harmful-content detection, support-ticket routing, urgency scoring, and combined escalation logic; the documentation also demonstrates thresholds and human review for uncertain cases. It positions Decision Models for classification, routing, triage, moderation, guardrails, agent tool-call approval, and model routing. It recommends conventional language models for generation, conversation, summarization, open-ended questions, complex reasoning, and code generation instead.