TypeSafe AI System One and Jev — Mission

What was and was not tested

This course was written from TypeSafe AI’s official documentation, its API reference, model page, and the official Jev launch post, read on 2026-09-20. The offline lab in code/typesafe-ai-system-one ran locally with Python 3.14.2: 19 tests passed, the mock route refused dispatch because explicit authority was absent, and a 12-case benchmark plan completed without network access.

No credential was read, no console session was opened, and no TypeSafe API request was made. Every live result, latency, token count, calibration claim on our data, and repository-specific benefit remains to verify.

Why I am learning this

Coding agents are good at producing text, plans and patches. Our repositories also contain smaller decisions that software must consume repeatedly: classify a Gaia candidate, score a GA search result, flag a Demerzel governance record, route an IX experiment, or identify a TARS grammar ambiguity. An LLM can answer those questions in JSON, but the surrounding program still has to parse it, reject invented values, measure uncertainty and keep authority out of prose.

Jev takes a narrower approach. It accepts a state plus typed questions and returns closed, structured answers. The model chooses among values we defined; our code still owns thresholds, composition and effects. That makes it interesting for high-volume judgment inside workflows, but it does not make a decision correct merely because its shape is valid.

The safety thesis

This course uses one rule throughout:

A model may estimate; deterministic code validates, gates and acts.

The distinction matters for every repository in our ecosystem:

• a valid Choice can still choose the wrong option;
• a high confidence is a model signal, not authority;
• a moving alias can change behavior without a code change;
• low cost can make a bad decision cheaper and more frequent;
• a provider response must never mint a Gaia grant, merge a pull request, or mutate governance by itself.

What you will build

The course’s Python lab has two paths:

1. Mock first: replay a saved response, validate its closed shapes and probability distributions, then apply the real routing policy with no network and no key.
2. One optional live probe: read TYPESAFE_API_KEY from the environment, estimate the input budget, make exactly one request, never retry automatically, never print the key, and record the request digest, concrete model, usage, latency and decision.

Outline

#	Lesson	Outcome
1	Decisions, not strings	Choose among Choice, Score and Noul without confusing types with truth
2	A reproducible, cost-bounded experiment	Run the offline baseline and understand the optional one-call probe
3	Use cases across our repositories	Select useful seams in Gaia, GA, Demerzel, IX and TARS while preserving authority
4	Token-cost benchmark	Test the 50% saving hypothesis against quality, retries and shifted cost
5	Confidence-gate stress test	Measure false support and review load on a deliberately wrong offline fixture
—	Journal	Measured facts, open questions and live work still to verify

Prerequisites

• Python 3.10 or later; the lab uses only the standard library.
• JSON and ordinary branching logic.
• An optional TypeSafe account and API key only for the live probe. Put the key in TYPESAFE_API_KEY; never paste it into a lesson, source file, terminal transcript or chat.

Primary resources

• Introduction and quick start
• Primitives, confidence, and patterns
• Models and pricing and the HTTP API reference
• Introducing System One Models & Jev, including the authors’ own caveats about early access and evaluation bias