Home / Writing / Jev: A Decision Model, Not Another LLM
Technology · · 2 min read

Jev: A Decision Model, Not Another LLM

TypeSafe AI's new Jev model returns structured decisions with confidence scores instead of text. Here is what is publicly claimed, what is unproven, and where it could sit in an e-commerce architecture.

A single glowing node splitting into three routes of different thickness, each ending in a probability bar

Jev is different from RAG and CAG in an important way: it is not a pattern for feeding a language model, it is a different kind of model. TypeSafe AI, a startup that came out of stealth in mid-September 2026, describes it as a “System One” model. Instead of writing text, you give it a question with a fixed set of answer options, and it returns a decision with a calibrated probability and a confidence signal.

Per the company and early press coverage, responses arrive in a fraction of a second at a tiny price per million tokens. Those figures are largely the vendor’s own claims, with one outside test on a single task, so treat them as promising rather than proven. It is proprietary, and reports flag real limits: you must define the questions, options and thresholds yourself, the model gives confidence rather than reasoning, and the provider has little production history.

Where it could fit in commerce

An e-commerce platform makes thousands of small, boring decisions. Is this support ticket about delivery, refund or sizing? Is this review spam? Is this product photo in the right category? Is this order worth a fraud review? Using a large language model for each one is slow and expensive. A decision model is built for that layer.

The proposed shape

Place Jev as a fast triage layer in front of the rest of the stack. Every incoming request gets classified first. Simple, high-confidence cases route straight to a rule or template. Ambiguous ones escalate to an LLM, backed by the CAG pack for stable knowledge and the RAG gateway for live facts. Low-confidence or high-stakes decisions go to a human.

Architecture diagram: Jev classifies each request and routes it to a rule or template, to an LLM with CAG and RAG, or to a human, with a fallback model

That gives a tidy division of labour: Jev decides where a request goes, CAG supplies what the business always knows, RAG supplies what changed today, and the LLM writes the answer.

Before you commit

Pilot it on one narrow task and measure it against your own labelled data. Keep a fallback model ready, because a single young vendor is a dependency risk. And do not use it where you must explain a decision to a regulator or customer.

Jev TypeSafe AI AI models e-commerce architecture solution architecture
Share