TypeSafe's Jev gives developers a calibrated probability instead of text output

Jev cannot return an answer that is not already in the set of outputs the caller defined before the call, and that is what the claim that it cannot hallucinate comes down to ^[1][19]. It can still put most of its weight on the wrong label. Armin Ronacher, the CTO of Earendil, which builds the open-source model harness Pi, told TechCrunch that "At the end of the day, it delegates the hallucination problem a little bit to the user" ^[10].

That delegation is a line of code somebody on the team owns. Ronacher described the judgement it asks for: "The user has to say, okay, if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it's 95%, sure, then I can do something with it" ^[11]. Nikhil Mudholkar, the CTO of Bryo AI, said of Jev that "it is the only one that hands back a real probability which makes it ideal for automating workflows!!" ^[9].

Output tokens are free and input tokens are metered by the billion ^[4]. That unit is a thousand times the million-token one ^[18]. TechCrunch did not publish a price per billion input tokens, so a buyer cannot turn Mudholkar's cost ratio into a line item without asking the company.

Almeida helped build ChatGPT at OpenAI and then helped invent reinforcement learning from human feedback ^[2]. "We have lightning in a bottle, and yet it is not useful," he told TechCrunch ^[13]. His diagnosis was that the field aimed at the wrong output: "We have been super good at human language for four years, but it's not useful for automation because computers speak a different language" ^[14]. He says Jev is trained exclusively on synthetic data, using a technique he calls "reinforcement learning from calibrated decisions" ^[15]. He was tight-lipped about the architecture, TechCrunch reported, and whether Jev is built on top of an open-weight LLM is unconfirmed ^[16].

Jev also gets used alongside LLMs. Almeida sees customers pointing it at LLM agent traces to catch jailbreaks, because using agents to monitor agents gets expensive fast ^[17]. Ronacher named model routing, where predicting which model a workload needs would be useful but too costly to do with an LLM ^[12].

Does the set of possible answers exist before the call is made? And does code branch on the answer, or does a person read it? Where the answers are enumerable and code branches, this is a candidate for the swap. The cost is that your team writes and defends the threshold instead of paying a vendor to hide it inside a model. If the answers are enumerable but a person reads them, the probability buys less, because the person is already the one judging. And when the output is a sentence somebody has to read, an LLM is still the thing that writes it. Both developers who reported numbers were replacing an LLM inside a classifier, the one job where the output set is written down in advance ^[6][8].