System One Models with Jev
Build software that asks a model for typed, probabilistic decisions instead of text, and knows when to trust the answer.
3 modules · 8 lessons~16 hoursUpdated 2026-10-31
- Lessons8
- Quizzes8
- Sources53
- Of study~16h
About this course
About this course
A practical course on System One models, the class of decision models introduced by TypeSafe AI with Jev. You learn what these models are and how they differ from LLMs and reasoning models, what calibrated probabilities mean and how to measure them, and how to use Jev's Choice, Score and Noul primitives well. The final module turns that into systems: routing on confidence, combining Jev with LLMs, and evaluating it honestly on your own data. Exercises run offline or on your Claude subscription through a local stand-in, and on the real Jev API once you have early access.
About the running example
The examples follow a fictional property-law (conveyancing) firm in England and Wales, ending with triage of the firm's inbox. It's a case study, chosen because its documents are long, structured and high-stakes. You don't need a legal background, and the same patterns apply to contracts, insurance claims, invoices, support tickets or patient records. In the lessons, a matter is one client's case and a solicitor is the lawyer who signs work off.
What you'll learn.
- 1Explain how a System One model differs from an LLM and a reasoning model, and decide which one a task needs
- 2Measure calibration with reliability diagrams, ECE and Brier score, and compute Jev's confidence values by hand
- 3Write Choice, Score and Noul questions with clear criteria over well-structured state, and avoid Jev's known failure modes
- 4Design confidence-gated workflows that act, confirm or escalate according to the stakes of each decision
- 5Combine Jev with LLMs in guardrails, filters and cascades, and evaluate the result on your own labelled data
3 modules, 8 lessons.
9h 15m of lessons
Module 1 · 2 lessons · 2h 30m
Decisions, not text
What a System One model is, why it is trained for calibrated probabilities instead of human preference, and what those probabilities mean.
Module 2 · 3 lessons · 3h 30m
The Jev API
Sending state and questions, the three question primitives, and how to structure state and decompose judgments so the answers are reliable.
Module 3 · 3 lessons · 3h 15m
Building systems
Turning probabilities into actions: confidence routing, hybrid workflows with LLMs, and evaluating Jev on your own data.
Triage a busy inbox with Jev
Build a confidence-gated triage pipeline for a synthetic business inbox, with typed Jev questions, stakes-based routing, an LLM-drafted reply behind a Jev guardrail, and an evaluation report that says when to trust it.