jev-harness
Research proposal-review contract: an LLM proposes one action, Jev answers four narrow questions, code produces evidence, applies nothing.
Overview
This jev-harness, from the independent TypeSafeAI community organisation (which the README distinguishes from the official TypeSafe team), is a research-stage proposal-review contract. An LLM proposes one action, Jev answers four narrow questions about it, and code produces evidence for a host to consider. Nothing in it applies a patch, executes proposed code or grants permission, and every step leaves a receipt.
Offline checks need no key: a bench command runs 25 synthetic proposal-review fixtures through validation alone and with a labelled mock transport, with the README stressing that the resulting totals are scripted mock values, not measurements of Jev. A live follow-up records a confidence-only catch and unexpected good holds and explicitly does not establish calibration. It uses a fixed 0.8 threshold, requires Node 22+ and pnpm, and is source-only. An optional Next.js demo starts offline.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Narrow review questions about a proposed action (four per proposal)(inferred) | — |
How it works
- LLM proposes one action
- Jev answers four narrow review questions
- Code validates and produces evidence with a receipt
- Host decides; the harness applies nothing
Strengths
- Clear separation of proposal, judgement and authority
- Honest labelling of mock versus live results
- Offline verification suite
Caveats
- Research stage; calibration not established
- Community project, not official
- Source-only, not published
Getting started
`git clone https://github.com/TypeSafeAI/jev-harness.git && cd jev-harness && pnpm install --frozen-lockfile && pnpm verify && pnpm bench:review`.At a glance
- Type
- Library
- Status
- research
- Language
- TypeScript
- Runtime
- Node 22+ library with optional Next.js demo
- Has code
- yes
- Has tests
- yes
- License
- MIT
- Stars
- 30
Use cases and integrations
How curated lists describe it
A custom coding harness for TypeSafe AI's Jev: an LLM proposes, Jev answers narrow questions, code decides, every step leaves a receipt.
wh000wh000/awesome-jev-liveA custom coding harness for TypeSafe AI's Jev: an LLM proposes, Jev answers narrow questions, code decides, every step leaves a receipt.
hellogumbo/awesome-jevResearch-stage proposal-review contract: an LLM proposes one action, four narrow questions are answered, and code produces evidence for a host to judge — it applies nothing itself. From the independent `TypeSafeAI` community organisation, which its README distinguishes from the official team.
MrJev/awesome-jev
Frequently asked questions
Is this from TypeSafe?
No, it is from an independent community organisation named TypeSafeAI.
Does it apply changes?
No, it only produces evidence for a host to judge.
Are the bench numbers Jev measurements?
No, offline totals use a scripted mock transport.
Related projects
GitHub Action that asks Jev three yes/no questions about a pull request and approves it only when all clear a confidence threshold.
ProjectbetaTypeScriptUnix-style tools where Jev learns a repo's decision norms and judges each node of a code lead as a bug, lead or clean.
ProjectbetaJavaScriptCoding-agent tools on Jev with public benchmarks: a strict command gate, a code-review router Action, and use cases.
MCP serverbetaTypeScriptCompiler that turns agent rules in CLAUDE.md and similar files into testable Jev programs: typed questions plus a coded verdict.
ProjectbetaTypeScriptZero-dependency Python MCP server where Jev gates shell commands, audits patches and resolves decisions, with SQLite caching.
MCP serverbetaPythonAdvisory Jev sidecar for Pi: four pre-flight hazard checks before bash/write/edit and four diff-quality checks after edits.
PluginbetaTypeScript