el-jev
Decision sidecar and Copilot CLI pre-turn hook that mimics Jev choice, noul and score using Azure Foundry Cohere Rerank.
Overview
el-jev is a fast typed-decision sidecar that recreates Jev's choice, noul and score primitives on Azure AI Foundry, converting Cohere Rerank relevance scores into decision probabilities, with keyless Entra ID authentication. It runs as a local daemon reachable from a GitHub Copilot CLI pre-turn hook, a CLI, MCP or custom applications.
It is designed for auditable abstention: by default non-trivial decisions stay advisory, returning selected or needs_review, until calibration has been fitted on labelled data. The README is careful about how it differs from Jev: Jev is a purpose-built model inside application-owned control flow, while el-jev's Copilot hook only adds guidance to the prompt, and Copilot still runs every turn and may follow or override it. It suits bounded routing, screening, gating and triage experiments.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option best fits this turn's intent (routing annotation)?(inferred) | — |
| Yes / no | Does the input satisfy this bounded condition?(inferred) | yes · no |
| Score | Where does the input fall on a defined scale?(inferred) | — |
How it works
- Copilot CLI hook or app sends a decision request to the local daemon
- Daemon scores options with Foundry Cohere Rerank
- Scores become probabilities; uncalibrated decisions return needs_review
- Caller (or Copilot prompt guidance) acts on the result
Strengths
- Keyless Entra ID authentication
- Abstention-first by default
- Reusable beyond Copilot via CLI/MCP
Caveats
- Not Jev; reranker scores are only converted to probabilities
- Copilot hook is advisory and can be ignored
- Requires Azure AI Foundry
Getting started
Deploy Cohere Rerank on Azure AI Foundry, run the el-jev daemon locally and install the Copilot CLI hook per the README.At a glance
- Type
- Hook
- Status
- prototype
- Language
- Python
- Runtime
- local sidecar daemon + Copilot CLI hook
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
High-speed System One decision sidecar & Copilot CLI pre-turn hook powered by Azure AI Foundry Cohere with keyless Entra ID auth.
logicrw/awesome-jev-projectsHigh-speed System One decision sidecar & Copilot CLI pre-turn hook powered by Azure AI Foundry Cohere with keyless Entra ID auth.
hellogumbo/awesome-jev
Frequently asked questions
Does el-jev use Jev?
No, it mimics Jev's primitives using Cohere Rerank on Azure Foundry.
Does it control Copilot?
No, the hook only adds advisory guidance; Copilot decides.
Why abstain by default?
Decisions stay advisory until calibration is fitted on labelled data.
Related projects
Decision-forensics library that explains a Jev answer by ablating parts of the input and measuring how far the probability moves.
ProjectbetaPythonAdapter that turns existing LLMs into Jev-style decision functions with option scores, optional calibration and review thresholds.
ProjectprototypePythonExact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch code showing confidence cascades let a 0.8B prefill-only decision model approach a 4B model's accuracy at a fraction of the cost.
ProjectresearchPythonResearch framework (DART) testing whether Jev's probabilities are a reliable signal to answer, escalate or abstain.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPython