Jev can't return malformed JSON. It can still pick the wrong allowed answer

When TypeSafe says zero hallucination, it means Jev cannot return an off-schema answer or a malformed tool call, because its possible outputs are fixed before the call ^[7]. The company says schema matching is guaranteed and type errors are mathematically impossible ^[25]. Jev can still pick the wrong allowed answer ^[7]. The only signal a developer gets for that is the confidence estimate returned alongside the decision ^[5].

The speed comes from the same restriction. A chat model generates tokens sequentially even when the API pins its final response to a JSON schema ^[6]. TypeSafe says Jev samples multiple typed outputs in parallel and gives up open-ended string generation entirely ^[6]. Behind that is a new architecture, a parallel sampler, and a training method the company calls Reinforcement Learning for Calibrated Decisions ^[9]. The founders spent roughly two years on the architecture, sampler and training process ^[8].

The Pong recording and the marketing multiples measure different things. Against Jev's 227 millisecond average, Claude Haiku 4.5 at 2.5 seconds is about 11 times slower, Gemini 3.8 Flash at 3.2 seconds about 14 times, and GPT-5.6 Sol at 3.5 seconds about 15 times ^[27]. The 47 decisions in the first 12 seconds work out to 255 milliseconds each, close to that reported average ^[28]. TypeSafe's own claim is 40 to 200 times faster on similarly intelligent System One-shaped queries, with a 193.6x speed example on its homepage ^[11][12].

Those larger numbers come from workflow evaluations that are harder to read across. They use identical compute graphs and score outputs against the average probabilities from GPT-6 Astra and Fable 5.1, not against ground-truth labels ^[13]. Four workflows are published, and the company's own model-capabilities team built them ^[14]. The LLM baseline runs through TypeSafe's System One wrapper, which constrains conventional models to compatible structured decisions and may affect their speed and cost ^[15]. Most tests ran from laptops on the US West Coast, where the service is based ^[16]. TypeSafe says the homepage figures likely sit near the upper end of real-world results and that long-term pricing sustainability remains unproven ^[12].

The demo conditions are documented: structured-output requests at temperature zero, reasoning disabled, no retries, one shared Vercel AI Gateway key, and latency measured on the server-side model call ^[21]. O'Riordan published the source code and the recorded statistics ^[31]. Every lane also receives an intercept coordinate, so the spatial reasoning that a 2024 Atari-GPT study found multimodal language models struggled with as low-level game controllers never comes into play ^[23].

"People can't be the only consumers of intelligence," Almeida said in TypeSafe's launch announcement ^[3]. OpenAI credits him under "Foundational RLHF and InstructGPT work" in its GPT-4 contributor record ^[2], and he describes Jev in the September 15th launch post as a "frontier-intelligence function call" ^[4]. For any of this to transfer to another workload, the decision has to fit inside 255 choices, or tolerate the two-stage score-then-select pipeline larger choice sets require ^[17]. And the confidence numbers have to be calibrated against that workload's own labels. The published evaluations test something else ^[13].