TypeSafe launched Jev a few days ago. I joined the waitlist, and I had a key a couple of hours later.
Jev is not an LLM. You cannot chat to it. That is the entire point.
typesafe.ai
It will not talk to you
People will try to prompt it like ChatGPT. It will not write code, summarise a PDF, explain why it answered that way, or generate a single sentence. You send it a state plus typed questions. You get numbers your code can branch on. TypeSafe calls this a System One model: a frontier-intelligence function call.
If you want prose, you still want Claude, GPT, or Grok.
Why I actually wanted it
I have a private Laravel app for my personal finances. Not a product, not a bank. Just the place I keep the numbers so I can answer “can I buy this right now?” without opening six spreadsheets.
The obvious AI move was a chatbot. Before Jev, I bolted a local assistant onto it with the Laravel AI SDK and Ollama (gemma4:12b). Streaming replies, conversation memory, nested specialist agents, PHP tools over those same calculators.
What I actually wanted was not a conversation. It was: given what is left, is buying X right now a smart idea?
Chat was the wrong product.
Follow-ups had no useful context. I asked how much I had per day for the rest of the cycle, then asked “how long left”. It answered with a completely different countdown from elsewhere in the app. I meant days left in the pay cycle.
It also picked the wrong figure for “outgoings left this month”: the calendar-month total, not unpaid in this cycle. I even built a 100-question live eval to catch it.
And it was slow. Thinking mode made the UI look frozen. Title generation could stall for about a minute. I capped tool steps, turned thinking off, killed title gen, and still had a 12B decode plus extra hops for every specialist agent. I spent time trying to speed the model up (MLX, smaller Qwen) instead of changing the shape.
Then I used Jev’s playground, and it clicked. All I care about is getting answers to decisions.
Noul, Score, Choice
Every question is one of three primitives. You can mix them in a single request. They all see the same state, they run in parallel, and your code composes the answers.
Noul is yes/no as a probability from 0 to 1. Near 1 is a strong yes, near 0 is a strong no, near 0.5 is a coin flip. It is not “medium”, and there is no separate confidence field. In the buying screen I ask whether the purchase fits available cash, whether it would harm savings goals, and whether the timing is sensible. The cycle screen asks whether spending is on track until payday.
Score is an ordered rubric you write, 2 to 10 levels. Jev returns a score that can sit between two levels, a probability for each level, and confidence. Stretch of a purchase is Comfortable, Noticeable, or Reckless. Cycle risk is Fine, Watch, or Overspending.
Choice picks one of up to 255 labelled options. You get the winner, the full distribution, and confidence. For a purchase that is buy_now, wait, or dont_buy. For the cycle: keep_going, slow_down, or pause_discretionary.
TypeSafe’s advice is one snap judgment per question, then combine in code. The buying screen fans five questions out in one call:
'recommendation' => [
'type' => 'choice',
'instructions' => 'What should I do about buying this item right now?',
'criteria' => [
'buy_now' => 'Affordable and sensible to buy now.',
'wait' => 'Possible later. Wait for payday, clear bills, or a better moment.',
'dont_buy' => 'Should not buy given current cash, unpaid outgoings, or savings pressure.',
],
],The other four sit next to it: three Nouls (fits_available_cash, harms_savings_goals, timing_sensible) and a Score for stretch. One round trip. PHP decides what the UI actually says.
Wiring it in
Each decision is a small PHP class with state(), questions(), and compose(). Laravel builds JSON state from fact slices (spending cycle, cashflow, savings), posts that plus the question map to https://api.typesafe.ai/v1/systemone with jev-latest, and compose() turns the typed answers into a verdict the UI can show.
The hard rule that fell out of the chat eval: never ask Jev how much is left. Put available_after_outgoings in the state, then ask whether the purchase is prudent. Unpaid means unpaid in this pay cycle, never the calendar-month total.
If a Choice’s confidence is under 0.45, the composer refuses the optimistic path. buy_now becomes wait. keep_going becomes slow_down. The raw pound facts still sit on the verdict so I can see what it was looking at.
Live Pest evals sit behind TYPESAFE_EVAL_LIVE=1. A USB cable at £8 should come back buy_now or wait with an affordability noul above 0.55. A luxury car at £45k plus £800 a month should be dont_buy or wait with a fit noul under 0.45. Ground-truth tests assert the tool state matches the Support calculators, including the unpaid-vs-month-total trap chat fell into.
What I actually use it for
Should I buy this? Item, one-off, optional monthly. A cheap essential mid-cycle (that USB cable) is a high affordability noul and buy_now. A big discretionary buy when cash is tight is a timing noul that says wait, even if savings could cover it. Raiding a savings goal for a gadget is where harms_savings_goals is the useful number, not a chatbot talking you round.
Am I on track this pay cycle? No purchase, just burn vs remaining cash after unpaid bills. Takeaways stacking up mid-cycle should flip to slow_down or pause_discretionary before payday. The value is a typed verdict with a confidence I can threshold, not a paragraph of advice. The app still owns the policy.
It is stupidly cheap
Jev bills input tokens only at $0.042 per million. Output is free. You get $5 free credit a month. I have used it a lot: playground, live evals, consulting the tools. $0.0018.
A local 12B was slow and wrong. Asking a frontier LLM to emit JSON for a yes/no is paying generation prices for a gate. If your code currently asks an LLM “yes or no, pick a bucket, rate this”, Jev is a better tool. If you want it to write the email afterwards, that is still an LLM’s job. The docs and console are the place to start.
