TypeSafe launches Jev; Check Point reports flipped verdicts

TypeSafe AI’s Jev returns calibrated typed decisions, but Check Point says prompt injection flipped investment verdicts in most tested runs.

TypeSafe AI introduced Jev, which CEO Diogo Almeida—an InstructGPT coauthor—describes as a System One model for reliable software decisions rather than open-ended chat, trained with Reinforcement Learning for Calibrated Decisions instead of RLHF or rubric-based RLVR. Sources disagree on the category: Simon Willison calls it a new kind of LLM that returns numeric decisions, while a Hacker News headline frames the Check Point research as showing Jev is not a language model but still breaks like one. A MarkTechPost guide to typesafe-sdk 0.7.0 says Jev returns Choice, Score, and Noul probability primitives, citing $0.042 per million input tokens and free output tokens. Check Point reported that, in a fictional PonziCorp due-diligence test, an attacker controlling one section of an uploaded report flipped Jev to a low-risk invest verdict in every configuration. The strongest attacker succeeded in 25 of 27 runs, usually on the fourth turn, at about 50 cents each; untrusted labels and anti-injection instructions barely helped, while reasoning effort improved resilience on comparison models.

• TypeSafe AI CEO Diogo Almeida, an InstructGPT coauthor, describes Jev as a System One model for fast calibrated software decisions, trained with RLCD rather than RLHF or RLVR, and prioritizes intelligence per dollar over public benchmarks.
• Simon Willison calls Jev a new LLM category that returns floating-point answers for yes/no questions, ratings, and confidence; a Hacker News headline on the Check Point work says Jev is not a language model.
• A MarkTechPost guide to typesafe-sdk 0.7.0 says Jev returns Choice, Score, and Noul (yes/no probability) primitives instead of generated text.
• That guide cites a price of $0.042 per million input tokens and free output tokens, with patterns such as Pydantic models, asyncio fan-out, retries, and batching.
• Check Point tested Jev in a due-diligence assistant reviewing fictional PonziCorp and reported that control of one uploaded-report section flipped the verdict to low risk and invest in every configuration.
• The strongest attacker succeeded in 25 of 27 runs, typically on the fourth turn, at about $0.50 per break.
• Untrusted-document labels and anti-injection instructions made little difference; reasoning effort was the strongest defense on comparison models.
• Latent Space and Willison published on 2026-09-21; MarkTechPost and the Check Point/HN report followed on 2026-09-24.