Decision models: cheaper judges and faster guardrails @arizeai
Decision models: cheaper judges and faster guardrails  @arizeai
Uploaded November 2026 | Updated November 2026, 3 hours ago
What does it cost to run an LLM judge on every span?

​Decision models are small, fast models built to pick from a set of choices. You pass in the input and the choices. They return each choice with a probability that it is the right one. They do not write text. They run faster than an LLM, and they are cheap enough to run on every span.

​In this session we will cover:
- ​What decision models are, and how the input and output work
- ​A short demo of a decision model in action
​- Using them as eval judges
​- How they compare with an LLM on cost and speed
​- Using them as real-time guardrails
​- Where they fit, and where an LLM is still the right call
​
Who this is for:
AI engineers building and evaluating AI agents.

​What you'll leave with:
How to read a decision model's probabilities, and how to use one as an eval judge and a real-time guardrail.
Decision models: cheaper judges and faster guardrailsBuilding the Next Generation of AI Agents: How We Built AlyxHow to Debug AI Agents in Production | An Intro to Arize AXHow OpenAI Processes Millions of User Feedback Data Points | Rise of the AI Engineer | Stuart SyWhy Youre Using LLMs Completely WrongAgent skills demo: Instrument, eval, and fix your AI in minutesHomework 1 and 2 for AI Evals Course: System Prompt Engineering and Error AnalysisIntroducing Signal: The AI Agent That Fixes Your AI Agents | From Production Traces to Pull RequestsArize AX Demo 2026: Build Self-Improving Agents with SignalAI Agent Mastery Certification Course: Lab 3 – Agent ArchitecturesAI Agent Mastery Certification Course: Lab 5 – RAG for AgentsWhy Software Engineering Offices Are Quiet Now
Arize AI |

Decision models: cheaper judges and faster guardrails

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER