Lesson 06 · Jev 101
The first ten answers are TypeSafe’s homepage FAQ, copied verbatim from the expanded accordion on typesafe.ai (browser capture, Sep 20, 2026). Extra 101 questions follow — those are this guide’s wording, not the homepage accordion.
Official source:typesafe.ai homepage FAQ ↗
Verbatim · attributed. Title Case as on the TypeSafe homepage.
System One Models are a new class of AI model built for decisions inside software. Jev is TypeSafe’s first public System One Model, optimized for automation. Send Jev structured questions and get typed decisions with probabilities and confidence that your software can act on.
Jev is neither small nor an LLM, hence being off the intelligence Pareto curve.
Jev’s efficiency comes from optimizing for a different task. It’s built for structured decisions inside software, with an interface and training approach designed for that purpose. It understands language, but doesn’t generate free-form text or function as a chatbot.
Valid JSON gives software a format it can read. But forcing an LLM into that format can leave some of its intelligence on the table. System One Models are trained for structured decisions from the start, returning typed answers with calibrated probabilities. Your code can use those probabilities to decide when to act, request more information, or escalate.
Jev replaces sequential generation with parallel computation, answering multiple structured questions in a single request. That makes it fast and efficient enough to bring intelligence to everyday decisions in your software, including ones that never justified the cost before.
Fun fact: the jump from sequential to parallel is similar to that made by the transformer over RNNs (the breakthrough underlying today’s AI revolution).
Nobody has actually asked us because nobody thinks this is possible... but yes we can. If you have a use case that needs a speedier Jev, contact us at sales@typesafe.ai and tell us more.
We can serve Jev profitably at our current prices. Our goal is to make intelligence more affordable over time as we improve the technology.
Jev is designed for common-sense judgments: classifying content, routing requests, scoring responses, and evaluating information. You can combine these decisions within your code to form larger workflows. Some tasks requiring extended reasoning, such as complex mathematics or chess-like planning, may be better suited to large reasoning models.
Yes. Jev guarantees the shape of its answers, not that every decision is correct. If you provide a list of categories, it can’t invent a category outside that list, but it can choose the wrong one. Uncertainty is a feature! You can use Jev’s provided probabilities and confidence to set the threshold for when your software acts autonomously and when it needs further review: higher for higher-stakes decisions, lower when errors are less costly.
Determinism means returning the same result for an identical input. This is less valuable than consistency. We define consistency as making similar decisions when the meaning stays similar, even if the wording changes. Jev is designed for consistency.
Join the waitlist! Jev is in its early days, and we’d love to hear how you’re using it, what you’re building, and any issues you run into. Join our Discord to ask questions, share feedback, and meme with our team.
Educational answers grounded in docs and the corpus. They are not the homepage accordion. New to Jev? Start at What is Jev?.
No. ChatGPT competes in chat and generation. Jev competes for typed decisions inside software. TypeSafe positions System One as complementary: keep an LLM for writing, coding, and open-ended work, and add Jev where you currently prompt a chat model to classify, route, score, or gate.
No. Official docs: “jev-1.13 is not trained to generate text.” Pair the models: Jev decides “needs review” or “send this”; ChatGPT (or any LLM) writes the comment, email, or patch.
They are Jev’s three question types. Choice picks one option from labels you define. Score places the state on an ordered rubric you define. Noul returns P(yes) for a yes/no question. Multiple questions can be evaluated together in one request (company: independently, in parallel).
Reinforcement Learning for Calibrated Decisions. TypeSafe’s stated training algorithm for calibrated probabilities on System One tasks. Contrast with RLHF, which trains models to produce answers humans prefer in chat. Different objective, different product behavior.
Official launch FAQ: after William Stanley Jevons / Jevons paradox — cheaper, more efficient intelligence should increase demand, not just save cost. Prefer that official etymology over community “unknown” notes.
The official homepage still shows Join Waitlist as of this guide’s compile date. Docs also describe creating an API key in the TypeSafe console once you have access. Vercel AI Gateway lists typesafe-ai/jev as a platform path — label that as a platform option, not TypeSafe’s default. This 101 does not run its own signup form.
Yes. Official docs cite a Python package (typesafe-sdk, Python ≥ 3.10) and a JavaScript package (@typesafe-ai/sdk). There is also an evaluate path via Vercel AI Gateway. Pin snippets from official Docs when you publish — APIs move.
Official quickstart: POST https://api.typesafe.ai/v1/systemone with a Bearer TYPESAFE_API_KEY. Confirm the current request body in Docs rather than freezing a sample here.
No public weights. It is a managed API. Community “reproductions” are not TypeSafe.
It accepts CJK, but official models.md says English is the primary training language and other languages, including CJK scripts, are handled but not equally well. Test on your distribution and watch confidence.
Not yet. Official models docs: text / JSON / array of text only. No image, audio, or video understanding. Keep multimodal work on an LLM.
TypeSafe’s type-safety claim: it cannot invent outside your criteria. It can still pick the wrong allowed option. Include an “other” label if you need an escape hatch.
TypeSafe argues prompted confidence from RLHF chat models is often overconfident and inconsistent. RLCD targets calibration so thresholds in product code mean something you can design around. Validate on your labeled data either way.
No. Replace specific decision calls you currently send to an LLM. Keep the LLM for writing, coding, tool-using agents, and explanations. Litmus test: can you enumerate the valid answers before you see the case?
Official models docs: Jev is not fine-tuned or LoRA-adapted with customer data. Same weights for all accounts. Customize via state, instructions / criteria, and composition in code.
Published figures: 250,000 tokens/sec and 1,200 requests per minute, dynamically adjusting. Contact sales@typesafe.ai for higher enterprise limits.
Yes. Official docs cite jev-1.13.0 as the current ID, with aliases jev-latest and jev-preview. Aliases can move. Pin the version you tuned thresholds against.
Official API concepts: no. For Noul, the probability is the uncertainty signal. Choice and Score expose confidence derived from the distribution shape.
No. Cyera SafeType is a ChatGPT privacy browser extension that flags or anonymizes sensitive data. TypeSafe AI is a San Francisco lab whose product is Jev / System One Models. Different company, different category. This 101 is only about TypeSafe AI.
Official team page: Diogo Almeida (CEO), Erik Gafni (CTO), Sasha Sheng (COO). Company materials describe Almeida as a co-inventor of RLHF / InstructGPT, previously at Google Brain. Confirm bios on typesafe.ai/team before quoting them elsewhere.
Business Wire (press): about $40M seed led by DCVC; the company emerged from stealth around September 15, 2026. Press-attributed, not a TypeSafe blog number.
TypeSafe’s homepage workflow-eval numbers. The launch blog says those multiples are on the higher end of real-world gains and points to evals.typesafe.ai for methodology. Reference labels: average of GPT-6 Astra and Claude Fable 5.1 high thinking. This 101 does not independently verify them.
Yes, as a decision primitive. TypeSafe publishes an agent skill (typesafe-ai/skills). Agents still generate code via LLMs. Jev is the classify / route / score / gate step, not the coder.
Do not act. Escalate to a human or a larger generative model. Official docs describe confidence patterns for automate-versus-escalate policy. Log outcomes and recalibrate.
Yes. Official theme: speculative fan-out. Ask many independent questions in one request — often cheaper and faster than serial chat calls. Code selects which answers to use.
Official jaggedness page: Jev is not a calculator and does not count reliably. Extract semantically, then compute in code. Same for “which date is earlier.”
Official docs: state is treated as data, not as a hostile user by default, and can be steered. Use precise criteria, tests, and your own injection / abuse defenses around automation. Improvements are planned; do not assume a built-in firewall.
Docs mention ZDR for enterprise customers (Legal). Vercel AI Gateway has per-request ZDR options — that is a platform feature, not TypeSafe’s default. Confirm current terms with TypeSafe.
Different failure modes, cost shape, calibration, and parallel multi-question decisions. TypeSafe’s line: System One is not JSON mode.
Company niche: runtime-defined labels plus language understanding without per-task training or customer LoRA.
It is early access. Attribute metrics as company claims. Read TypeSafe’s jaggedness page and evals caveats before you automate high stakes.
TypeSafe publishes an anti-benchmaxxing policy and refuses standard bench tables in model releases. Use private evals plus the workflow-evals site.
No. It embeds as a primitive. Code owns control flow. Official docs: AI-powered software, not agents.
Published millisecond bands are from a West Coast service. Measure from your region. Do not promise homepage ms worldwide.