Choice
Pick from a known set. Routing, classification, next action.
a small intelligence layer for software
JevLM turns messy state into typed, probabilistic decisions your code can actually use. Route. Score. Verify. Escalate when the model knows it should.
“We were billed twice for March. Please refund the duplicate before Friday or we will cancel our plan.”
THE MISSING LAYER
Today, a developer asks a text model for a classification, waits for prose, parses it, validates it, and still has no honest measure of whether the answer is safe to act on.
JevLM is designed at the other end of that pipeline. Possible outputs are declared up front. Every answer carries a distribution. Your code decides what happens next.
Explore the primitivesTHREE BUILDING BLOCKS
One call. Independent questions.
No generated strings to parse.
Pick from a known set. Routing, classification, next action.
Grade a state against an ordered rubric. Keep the full distribution.
Ask whether a statement is true. Get a probability, not a shrug.
BUILT FOR THE REAL PIPELINE
A decision can be fast
and still be accountable.
JevLM gives production systems a place to pause, measure confidence, and hand off gracefully when the evidence is thin.
Choices, rubrics, and truth checks are defined before inference.
Return the answer and the uncertainty your policy needs.
Set a threshold for automation. Route the rest to a person.
THE THESIS
Not another chatbot. Not a JSON wrapper around a text generator. A fast, typed decision primitive for the thousands of small judgments that make products feel intelligent.
EARLY ACCESS / 2026
Tell us what your software needs to decide. We’re collecting the first workflows for the private build.