Jev-Engineering

Open research and tooling around JEV: structured decisions instead of free text.

JEV is a model that answers small structured questions (for example, which of five known actions fits a situation) instead of generating free text.

What we do

We build open tooling around that idea: finding where a decision point in a codebase might benefit, compiling typed decision questions, and researching parallel typed decisions and calibration.

Repositories

Repository	Summary	License
jev-integration-evaluator	Scan a codebase for decision points where JEV might help, offline and read-only, and test them with evidence	MIT
TypeWright	Declare a decision, compile typed Jev questions, measure them, and ship a JSON program	MIT
layev	Open research implementation of parallel typed decisions and calibration inspired by Kev and Laya	Apache-2.0
jraphyte	TRACE-GC evidence-bound graph synthesis, GraphRAG, and first-page research validation	MIT

Limits

• layev is a research candidate, not a validated Jev replacement. Its CPU fixture tests pass; native Qwen/CUDA checks and Jev-relative quality are still open.
• jraphyte is a reference implementation with explicit deployment gates, not a qualified production service. Its demo uses fabricated observations and no provider.
• jev-integration-evaluator is read-only and offline by default.
• TypeWright is alpha: its no-key demo and core tests pass, but no live provider study has been completed and no measured accuracy gain is claimed.
• Live provider results are not claimed here; each README states what was and was not exercised.

Contributing · Support · Security · Steward: Complete Tech LLC