pilot-typesafeai-jev
Pilot that puts Jev into LangGraph routing and Deep Agents guardrails and claim verification, with executed notebooks and caveats.
Overview
pilot-typesafeai-jev is a pilot project that integrates TypeSafe's Jev into LangGraph and Deep Agents, running chat models on ChatGPT, NVIDIA NIM or LM Studio, and records what works and what does not. Code owns the workflow, Jev makes the fast judgment calls, and a chat model still writes replies.
Case 01 uses Jev as a LangGraph router: one request asks three questions about a message and plain code routes it to answer (the only route spending chat-model tokens), escalate (urgent), review (unclear) or refuse (injection). Case 02 adds a guardrail middleware in Deep Agents that screens messages before the agent starts and a verify_claim tool returning a verdict with confidence. Results come from executed notebooks repeated several times on LM Studio with gemma-4-e4b, with verification documents. The README states it is not a benchmark of Jev accuracy, thresholds are untuned, it is not production code, and the ChatGPT subscription provider is unofficial. Documentation is in four languages.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Triage: should this message be answered, escalated, reviewed or refused? | answer · escalate · review · refuse |
| Yes / no | Does this message contain a prompt injection? | yes · no |
| Yes / no | Is this claim verified? | yes · no |
How it works
- Message enters a LangGraph triage node.
- One Jev call answers three questions.
- Code routes; only answer uses the chat model.
- In Deep Agents, middleware screens input and verify_claim checks outputs.
Strengths
- Clear separation of code, Jev and chat model roles.
- Executed notebooks with published caveats.
- Multilingual docs.
Caveats
- Not a benchmark; thresholds untuned.
- Not production code.
At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- LangGraph, Deep Agents, notebooks
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
👩🔬 Pilot of the decision model 'jev' from TypeSafe AI.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev do in the LangGraph case?
Answers triage questions so code can route to answer, escalate, review or refuse.
Is it production-ready?
No, it is a pilot with untuned thresholds.
Which chat models were used?
LM Studio with gemma-4-e4b for the recorded results.
Related projects
Terminal Q&A app where Jev sets response policy and reviews drafts while OpenAI only writes, orchestrated with LangGraph.
AppprototypePythonPython multi-agent runtime with native Jev support: agents ask typed questions via a decision client separate from the text model.
LibraryproductionPythonOpen-source agent guardrails that use Jev or LAYA to check prompts, retrieved content, tool calls and outputs for injection and risk.
ProjectbetaPythonjevctl: Node CLI for typed Jev judgments (verify claims, screen agent hijacks, classify) with exit codes for CI gating.
CLIbetaTypeScriptUnofficial Python adapter that brings Jev verification, classification and scoring into the Mellea framework as Requirements.
ProjectbetaPythonCompanion code comparing Google ADK and Microsoft Agent Framework for structured video timelines, with Jev as a shared per-segment QC gate.
LibraryprototypePython