# TypeSafe tests semantic lints before malformed questions reach Jev > Source: > Published: 2026-09-21 04:19:11+00:00 # TypeSafe tests semantic lints before malformed questions reach Jev **A console warning spotted by Jane Manchun Wong catches open-ended prompts sent to Jev's binary Noul primitive.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) ยท Published Primary source: [X - Jane Manchun Wong](https://x.com/wongmjane/status/2101822614218146061) ## Why it matters TypeSafe's reliability pitch depends on developers asking well-formed typed questions. Semantic lints catch a failure that JSON validation cannot: valid code expressing an invalid decision. TypeSafe AI, founded by [Diogo Almeida (@CompleteSkeptic)](https://x.com/CompleteSkeptic), Sasha Sheng and Erik Gafni, is testing a semantic linter that warns developers when a question does not fit the answer type they selected for Jev. Security researcher and reverse engineer [Jane Manchun Wong (@wongmjane)](https://x.com/wongmjane) surfaced the work [in a post on X](https://x.com/wongmjane/status/2101822614218146061) on September 20th. Wong is known for finding unreleased features by examining apps and web products, which makes the screenshot evidence of work in progress rather than confirmation of a general release. [https://x.com/wongmjane/status/2101822614218146061](https://x.com/wongmjane/status/2101822614218146061) The image shows TypeSafe's question editor accepting an object labeled `new_noul_1`. Its type is set to `noul`, while its instruction contains the open-ended prompt, "How are you from?" The editor places a warning beside the object: "Not a yes/no question." It then tells the developer to rewrite the instruction so it can be answered yes or no. A checkbox labeled "Enable TypeSafe semantic lints" appears inside the warning. That control suggests the checks are designed as an optional authoring layer in TypeSafe's console, catching errors before a request is sent to the model. ### A type-correct question can still be wrong TypeSafe released Jev in early access on September 15th after two years in stealth. Almeida, a former OpenAI researcher and co-author of the InstructGPT paper, has framed Jev as a machine-facing alternative to generative models: developers provide state and predefined questions, and Jev returns structured decisions rather than prose. Almeida built TypeSafe with Sheng, a former Meta and FAIR research engineer, and Gafni, a repeat founder who previously worked at Invitae and Freenome, according to [TypeSafe's team page](https://typesafe.ai/team). The San Francisco startup also announced a [$40 million seed round led by DCVC](https://www.dcvc.com/news-insights/typesafe-emerges-from-stealth-with-a-new-way-of-doing-ai/) alongside Jev's release. Jev's [public documentation](https://docs.typesafe.ai/introduction) defines three question primitives. `Choice` selects from a developer-supplied list, `Score` evaluates an input against an ordered rubric, and `Noul` returns a probability between zero and one for a yes-or-no proposition. Multiple questions can be evaluated in parallel against the same state. That structure prevents one class of model failure: Jev cannot invent a fourth answer when software has declared three valid choices. It does not prevent a developer from placing an open-ended instruction inside a binary question object. The JSON remains valid, and the output can remain type-correct, even though the question itself makes no sense for the selected primitive. The semantic linter addresses that gap. A conventional schema validator can check that `type` equals `noul` and that `instructions` contains a string. Determining whether the string describes a binary proposition requires interpreting its meaning. TypeSafe is bringing that interpretation into the development environment, closer to the point where the error is introduced. ### TypeSafe is building the tooling around its model The feature also sharpens TypeSafe's broader product strategy. Jev's value depends on developers decomposing work into narrow questions that software can safely act on. In its [launch post](https://typesafe.ai/blog/introducing-system-one-models-and-jev), TypeSafe said those typed decisions can be produced in 70 to 500 milliseconds and priced input at $0.042 per million tokens. Those remain company-reported figures, and TypeSafe acknowledges that its published workflow evaluations favor the model's intended workload. Early developer interest has centered on classification, routing, safety checks and other tasks where an application needs a bounded answer rather than generated text. [TechCrunch reported](https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/) that demand briefly exceeded TypeSafe's ability to serve API users after launch. Semantic lints fit the same thesis at a smaller scale. TypeSafe is treating AI instructions less like chat prompts and more like code that should be checked before execution. The warning Wong found does not make Jev's underlying judgment correct. It removes a preventable category of developer error before the model is asked to judge anything. That distinction matters for TypeSafe's pitch. Structured outputs constrain what a model can return; semantic validation constrains what developers can sensibly ask. Reliable automation requires both, and TypeSafe is starting to put the second constraint directly into its tools.