The hype lands as one sentence:
Everyone should switch to Jev. It’s much faster than the old LLMs.
That mixes three claims into one product pitch: speed, typed output, and a confidence number that sounds like certainty. None of those make Jev a substitute for a model that writes, argues, or thinks with you.
It’s a tool for a specific kind of act. Confused marketing skips the act.
System One vs System Two
Kahneman, Thinking, Fast and Slow:
- System 1: fast, automatic, cheap. A pattern comes in, a snap judgment goes out. “Trash.” “Urgent.”
- System 2: slow, deliberate, expensive. Holds contradiction, follows an argument, writes, checks.
TypeSafe named Jev a System One model on purpose: state in, typed decision out, no prose. They did not ship a “System Two model.” System 2 in their map is still LLMs, reasoning, and you.
The contract
State in. Closed options out: Choice, Score, or noul. (noul is their yes/no: one probability between 0 and 1 that the answer is yes, not a separate Boolean type.) Probability on those options. Code branches.
Whatever doesn’t fit the set is dropped on purpose. That isn’t a bug. It’s the deal.
The useful question is not which model is faster. It is: what is the objective, how will you measure it, and what must be true before any call runs? A rubric. Options another engineer would apply the same way. A threshold that reduces work without eating expensive mistakes.
Bad context in, irrelevant number out. Plausible output that hides that nobody decided what mattered. Only this time the prose is a decimal.
When the shape fits
Places where an agent already makes a yes/no or pick-one call, and often burns a frontier model on it today:
- Model routing: small vs frontier (needs a non-circular definition of “complex”)
- Tool gates: allow or block when risk depends on context, not only on an allowlist
- Intent triage: refund, bug, or deploy (volume plus a stable taxonomy)
- RAG scores: relevance per chunk
- Guardrails: policy that is semantic and narrow
- Escalate to human: auto vs review (only after calibration on your cases)
Pattern: Jev as a porter between agent steps. Not the system that reviews a PR against standards and history.
Some questions only look closed. “Continue the loop?” is a Choice on paper, but without signals you can observe (tests passing, diff size, error count), it is a System 2 judgment wearing a System 1 costume. Write the signals first, or keep that call out of Jev.
What confidence is not
The number feels like rigor. The rigor is in what was asked.
A closed set, a written rubric, and code that acts on the threshold are what make the number usable, every time you put Jev in a pipeline. Without those, you have a thermometer without a scale.
TypeSafe sells calibration through a training method it calls RLCD. Its launch post gives the name and the goal, not the method, and I couldn’t find calibration figures you can check on their side. So 0.90 only becomes serious after you measure it on your distribution of that decision, not because the API returned a decimal. Same rule as everywhere else with these tools: bad or thin context in, irrelevant confidence out.
auto @ 0.90 vs always human is a false choice. On expensive decisions there is a third path: the model recommends, you apply. Pass only the previous step’s confidence into “reliable enough?” and you are calibrating an echo.
Three questions
Before Jev goes into a pipeline, answer these.
- What decision no longer goes through an LLM, in one sentence?
- What are the exact options, or the schema?
- At
p = 0.62, what does code do? Not what you feel.
Without answers, you get prettier JSON. With them, you get a real System One queue.
The assessment line
Today, do I throw away the text and keep a repeatable yes/no?
If not, it isn’t Jev.
If yes, with a written rubric, a closed set, and calibration checked on your cases, then System One is the right shape. An LLM forced into JSON was always a workaround for that shape. Jev is honest about the shape. It is not a faster brain for every job.
Start from the objective. Work backward. Then pick the tool.