Key Takeaways
- Standard frontier models generate unstructured text strings, creating high latency, unpredictable output parsing, and steep output token costs.
- TypeSafe AI built Jev as a system-one decision model that takes unstructured text inputs and returns typed data primitives rather than raw prose.
- Jev charges only 4 cents per million input tokens and zero dollars for output tokens because the model emits almost no generated text.
- Engineering teams pair Jev with frontier LLMs and real-time voice APIs to handle pairwise pull-request categorization, local coding telemetry, and live user routing.
- Developers structure fast classification pipelines using Jev's Three Output Primitives for System-One Logic.
The Jev's Three Output Primitives for System-One Logic
Claire Vo contrasts traditional generative models with purpose-built decision engines. As Vo explains, “With the standard LLMs that you're used to working with, you are getting strings and generated text out. So you're getting text in, text out. With Jev, you're getting text in, type safe values out.” To run deterministic system-one logic without generation overhead, Jev relies on three distinct output primitives:
- Choice Primitive: Provide unstructured text and a predefined array of discrete choices. Jev evaluates the input and selects a single matching category (e.g., categorizing an email topic, selecting attire based on context, or routing an issue). Vo notes, “It will return a choice which means you can give it text, you can give it a list of choices and it will pick a choice.”
- Score Primitive: Evaluate unstructured input against an ordinal or numerical ranking scale (such as 1 to 5, or cosmetic/broken/blocking). Ideal for triaging severity, bug prioritization, or sentiment grading.
- Nule / Boolean Likelihood Primitive: Evaluate a binary proposition and return a probability score reflecting the likelihood that the answer to a specific question is 'yes' (e.g., assessing whether a user turn contains a feature request or if an email can be deleted). Vo describes this output: “It's a version of a boolean. It basically tells you what the likelihood that the answer to a question is yes.”
When This Works (and When It Doesn't)
Apply this framework whenever an application requires deterministic routing, filtering, scoring, or smart conditional classification over unstructured text without paying for generative token latency or output token fees. Because “Jev only charges you on input tokens because it barely outputs anything. And it is 4 cents per million input tokens,” high-throughput workflows like triaging every git commit or parsing live audio transcripts become economical.
This framework fails if your application requires creative prose, open-ended summarization, synthetic data generation, or reasoning across raw images. Vo points out that while Jev accepts text descriptions of images, it does not process raw pixel arrays directly. If your pipeline needs to draft an email response rather than simply route it, pass Jev's typed output downstream to a standard frontier LLM.
What to Do With This
Audit your LLM pipeline tomorrow and identify every prompt where you force a large model to return JSON with structured keys like category, priority, or is_churn_risk.
If you run an automated customer support intake, replace your expensive general-purpose LLM router with Jev's three primitives:
1. Use the Choice Primitive to assign each incoming ticket to a team by passing categories like ["billing", "bug_report", "feature_request", "sales"].
2. Apply the Score Primitive on a 1 to 5 scale to grade user urgency based on ticket text.
3. Run the Nule / Boolean Likelihood Primitive on the question "Does this user threaten to cancel their account immediately?"
Feed these three typed values directly into your database or routing logic. You eliminate output parsing failures, strip latency down to milliseconds, and cut routing costs to four cents per million tokens.