Unofficial Guides + Use Cases for TypeSafe’s System One / Jev. Facts from public docs. Not affiliated with TypeSafe.
What Jev is → vs chat models & code → Choice / Score / Noul → first small task.
01
Jev is TypeSafe’s System One decision model: send state + typed questions, get structured answers your code can branch on—illustrated with a support email, not a chat.
02
Code owns control flow and side effects; Jev owns narrow semantic judgments; LLMs write prose when you need text. Prefer workflows + gates over open agent loops.
03
Choice selects from a closed set; Score rates ordered levels; Noul returns P(yes). Mix them in one call; each question runs in parallel against the same state.
04
Shrink “AI handles support” to a reviewable label on desensitized mail. Define input, options, success checks, and what you refuse to automate.
Catalog of decision shapes — who, steps, expected effect, sources. Not chatbots.
Who: Teams wrapping Codex/Claude Code/Pi-style agents. Steps: On turn/tool boundary ask parallel questions — skill Choice, action allow/ask/deny, stuck/slop/done Nouls → harness steers or blocks. Expected effect: cheap semantic supervisor around the agent (community cluster).
Get the prompt →Who: Builders of long-running agents with persistent memory. Steps: On write, ask Noul / Choice questions for memory type, redundancy and links to stored memories; on read, ask which memory views to search, how much budget to spend, which candidates to keep and whether the evidence is enough; code enforces the limits and passes the selected evidence to the answer model. Expected effect: fewer language-model calls inside the memory loop.
Get the prompt →Who: Messaging power users. Steps: (1) Capture chat via OCR/a11y. (2) Closed Choice/Score/Noul on intent, danger, reply-now. (3) Code ranks; generative model drafts text only. (4) Human sends. Expected effect: judgment before generation; no auto-send.
Get the prompt →Who: Dev workflow. Steps: Diff summary → Choice allow/ask + Nouls → pre-commit. Expected effect: Local quality gate.
Get the prompt →Who: Multi-agent runners. Steps: Worker trace → Nouls complete/tested/off-track/stuck → supervisor acts. Expected effect: Architecture experiment — label as such.
Get the prompt →Who: Tooling / trading demos. Steps: Map NL to Choice function + closed-set args → code calls. Expected effect: Confidence-aware tool binding.
Get the prompt →TypeSafe + Gateway (official paths) and independent demos with clear boundaries.
official
Official TypeSafe Jev: docs, playground, console API keys, Python/JS SDKs. Returns structured Choice / Score / Noul answers — not chat.
integration
Call Jev through Vercel AI Gateway (`typesafe-ai/jev`) with the AI SDK evaluate path. Useful when you already route models through Gateway.
community
SemIf reproduces a Jev-style interface on open models. It is independent and does not reproduce TypeSafe training or hosted Jev accuracy. Never cite SemIf numbers as “Jev accuracy.”
community
Community demo that pairs perception with decision-style calls. Independent project — not an official TypeSafe product.
community
Independent extension demo in the broader Jev-adjacent ecosystem. Not affiliated with TypeSafe.
Prepare materials → support routing → RAG filtering → n8n → extract and validate.
hands-on
External note.com write-up by Takuma Yoshida (2026-10-29): on Kaggle Game Arena Bargaining, splitting negotiation principles into many small hosted TypeSafe Jev questions (then optionally moving calculable parts to Python) beat stuffing seven principles into one prompt.
hands-on
External DEV write-up by Olivier Miossec (posted 2026-10-09, edited 2026-10-25): the mental model behind hosted TypeSafe Jev—state vs questions, Score / Noul / Choice, confidence thresholds, and writing atomic questions—with PowerShell examples that apply in any language.
hands-on
External DEV write-up by Olivier Miossec (2026-10-01): use Doug Finke's Jev PowerShell module to gate Bicep What-If changes, review NSG rules, and triage Azure Policy linter findings with hosted TypeSafe Choice / Noul / Score questions.
Is Jev an LLM? What is Noul? OpenJev ≠ Jev. Confidence misuse.
Repeating launch-blog multiples as independent Jev 101 benchmarks.
Those figures are TypeSafe workflow-eval marketing on their launch blog—we do not re-run them here.
HTTP 401 Unauthorized from TypeSafe API or Gateway when calling System One / jev.
Unauthorized usually means missing/invalid TYPESAFE_API_KEY, wrong auth header, or calling the wrong host.
Treating every Choice/Score as infallible automation fuel.
Yes. Use confidence paths, human escalation, and domain evals. Honest uncertainty is a feature.