Hinge-Jev
Decision engine that combines Jev probability distributions with loss matrices to decide when a clarifying question is worth asking.
Overview
Hinge-Jev is a TypeScript decision engine and local visual studio for ambiguous automation requests, such as "clean up the old Atlas records" where it is unclear whether to archive or delete, or which environment is meant. It is for developers building agents that must decide between acting, asking, and deferring to a human.
Jev provides a joint Choice distribution over possible interpretations (worlds). Hinge then applies a developer-written loss matrix, stress-tests the distribution with a Total Variation radius to guard against miscalibration, and ranks candidate clarifying questions by value of information minus interruption cost. It asks a question only when the net value is positive, conditions the distribution on the answer, and otherwise proposes a bounded action or defers when risk exceeds budget. Each run can export a SHA-256 signed receipt for replay. An example mode works without an API key using authored distributions; a live mode calls Jev.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which interpretation (world) of the ambiguous request is intended, e.g. archive vs delete, staging vs production?(inferred) | — |
| Other | Is asking a clarifying question worth the interruption cost? | ask · act · defer to human |
How it works
- Receive an ambiguous natural-language request
- Jev returns a joint Choice distribution over worlds
- Compute expected and TV-stressed loss per action from a loss matrix
- Ask the highest-value clarification if net value > 0, then condition on the answer
- Otherwise act within risk budget or defer to human review
Strengths
- Makes the cost of acting on a wrong interpretation explicit and inspectable
- Worst-case stress test protects against overconfident probabilities
- Signed receipts make each decision replayable for governance
- Example mode runs locally with no key
Caveats
- Example mode uses authored fixture distributions, not live model output
- Loss matrices must be written by the developer for each domain
- Live mode needs TypeSafe early-access credentials
Getting started
Requires Node.js 22+: `git clone https://github.com/satiricalguru/Hinge-Jev.git && cd Hinge-Jev && npm install && npm run dev`, then open http://127.0.0.1:5173.At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- Node, browser studio
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
🎯 A Jev-powered decision engine and visual studio that prices uncertainty and asks the question that matters.
logicrw/awesome-jev-projects
Frequently asked questions
What does Hinge decide?
Whether to act, ask a clarifying question, or defer to a human, based on Jev's distribution over interpretations and your loss matrix.
How is it different from just taking Jev's top answer?
It weighs every possible interpretation by the cost of being wrong and stress-tests the probabilities before acting.
Do I need a TypeSafe key?
Not for example mode; live Jev inference needs TYPESAFE_API_KEY in .env.
Related projects
- SystemOneHarness★ 193
Agent harness that compiles an environment's finite actions into typed questions for Jev and gates each step by confidence.
CLIbetaPython Playbook for using Jev as a cheap gate before expensive agent steps: confidence policy, question recipes, a skill and a router.
ProjectprototypeMinimal agent loop where Jev selects each action and approves tool calls, a chat model fills arguments, and MCP tools execute.
MCP serverresearchPythonZero-dependency agent harness that sends each loop fork (act?, which tool, risk, done?) to a decision model and keeps the LLM for writing.
ServiceprototypeJavaScriptAgent skill that sends closed coding-agent judgments to Jev before asking the user, before risky commands and after failures.
ModelbetaPythonDeepSeek Harness preset where Jev answers an agent's open end-of-turn questions, steering it on only above confidence and safety thresholds.
Articleprototype