Can Jev Generate Text or Write Code?
Last checked · Independent guide, not affiliated with TypeSafe AI
No. Jev only returns typed answers you define in advance: a choice from a list, a score on your levels, or a yes/no probability. It cannot write replies, summaries, translations or code. When a task needs text, developers let an LLM write it and use Jev to decide, route or check around it.
Jev understands text but does not produce it. TypeSafe’s documentation states that System One models do not write replies, produce code or explain their reasoning. Every answer is a value from a space you defined in the request.
Why it cannot write
Section titled “Why it cannot write”Generating text means choosing the next word, then the next, thousands of times, where each choice depends on the last. Jev is built for the opposite: it evaluates all of your questions in one parallel pass and returns probabilities over fixed answers. That design is what makes it fast and cheap, and it is also why there is no text to return.
“Can I just make the answer space very large?” came up on Hacker News: for example, a Choice over every possible next character. TypeSafe addresses this in its Jev 1.13 notes: you can force string output by chaining choices, but it works badly and is very slow. A Choice is also capped at 255 options.
What about code?
Section titled “What about code?”The same applies. Jev cannot write a function, fix a bug or produce a diff. It can, however, make decisions about code that an LLM or a human wrote:
- Does this change touch files it should not?
- Which reviewer group should see this pull request?
- Does this error log look like a timeout, an auth failure or a crash?
- Is this generated test meaningful or trivial?
Early projects used Jev this way for code review checks and for routing coding tasks to the right model. The official agent skill for Claude Code and Codex exists so that coding agents write Jev calls into your software; see Jev with Claude Code.
Extraction without generation
Section titled “Extraction without generation”A common workaround for pulling a value out of text: generate the candidates another way, then let Jev choose.
- Use a regular expression or a parser to find candidate values (every date, email or amount in a document).
- Ask a Choice: which of these candidates is the invoice due date? Include a “none of these” option.
- Normalize the chosen value in code.
TypeSafe’s cookbooks use this pattern for dates, emails and amounts. It keeps the answer verbatim from the source, so nothing is invented.
Pairing Jev with an LLM
Section titled “Pairing Jev with an LLM”| Pattern | Who writes | Who decides |
|---|---|---|
| Browser agent (Jev Ultrafast) | A small LLM types text into fields | Jev picks the action and element |
| Model routing | The chosen LLM writes the answer | Jev picks which model gets the request |
| Guardrails | The LLM writes the reply | Jev checks the prompt and the reply before they pass |
| Drafting support replies | An LLM drafts | Jev scores tone, urgency and whether escalation is needed |
The division of labor keeps the expensive, slow model for the parts that truly need words.