Jev Decision Gateway

A gateway that sits in front of expensive generative LLMs. Application and agent code asks Jev the small, bounded, structured questions it is good at; a deterministic policy layer turns the typed, confidence-scored answers into an execution plan; and a real LLM is invoked only when policy says generation is needed.

Modern agents make many small structured decisions — should the agent keep going, which tool next, is this document relevant, does this output pass, should we escalate. Routing each one through a generative LLM pays generation-level latency and cost for what is really a bounded classification or a boolean gate.

From the README: one gateway.evaluate() call returns nextTool, shouldContinue, verifyOutput and modelTier; only when finished does one LLM call generate the final answer, at the tier Jev picked. The before picture used four LLM calls per iteration, three of which never needed generation.

No, and the project says so. Jev cannot write a paragraph, synthesize a summary, produce code or carry out open-ended multi-step reasoning. Every example and benchmark still calls a real LLMProvider for the generation half of the work; Jev only gates when and how generation happens.

Run npm install, then npm run build — the README stresses the build is not optional because packages point main/types at ./dist with no dev-mode source alias — then cp .env.example .env and fill in the TypeSafe API key.

The README says the flagship example (examples/agent-runtime-optimizer) and the benchmark (benchmark/ai-agent-scenario) implement exactly the gated shape. No benchmark results appear in the captured excerpt, so performance claims are unverified in this entry.

No. npx jev-gateway only works from inside a clone of the workspace; the package is marked private and unpublished, and the README warns it is not something installable from the npm registry.