jev-learnings
Field guide site on decision models built from two Jev projects: a code, decision model, LLM, human cascade from ELI5 to engineering.
This repository is mostly documentation or write-up rather than runnable code.
Overview
jev-learnings collects generalised findings (rates, AUCs and patterns, with no personal data) from two projects that used Jev as a cheap, fast judgement tier between deterministic code and an expensive LLM: an email first pass over a personal inbox, and an agent that plays the idle game Melvor Idle in a browser for days at a time.
The companion site, Decision Tier, runs from ELI5 through 101, 201 and 301 lessons to architecture pages on the cascade and on novelty and doubt. A four-page Melvor case study covers the testbed, why a game, the cascade as built, and 51 resolved escalations and where each fix landed. Evidence pages compare Jev, OpenAI's Decisions API, open-weight decision models and Claude Haiku 5.5, include a cost ladder with a calculator, and an October 2026 landscape refresh. A facilitated Claude Code tutorial builds a 'decision skeleton'.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this email need attention? (first-pass example)(inferred) | — |
| Choice | Which game action should the agent take next?(inferred) | — |
How it works
- Code handles certain cases
- Jev judges the uncertain middle
- Escalate doubtful or novel cases to an LLM
- Escalate further to a human; push fixes back down the cascade
Strengths
- Grounded in two long-running projects
- Graduated learning path
- Cost ladder and comparisons
Caveats
- Generalised findings only; raw data not published
- Mostly documentation
Getting started
Read https://robgruhl.github.io/jev-learnings/ or run `cd tutorial && claude` for the guided tutorial.At a glance
- Type
- Article
- Status
- production
- Language
- Python
- Runtime
- Static site and Claude Code tutorial
- Has code
- yes
- Stars
- 0
Use cases and integrations
How curated lists describe it
Decision Tier: a field guide to decision models (Jev and peers), from ELI5 to engineering a code → decision model → LLM → human cascade
wh000wh000/awesome-jev-live
Frequently asked questions
What is the cascade?
Code first, then a decision model, then an LLM, then a human, with fixes pushed down over time.
What projects is it based on?
An inbox first pass and a long-running Melvor Idle agent.
Does it compare alternatives?
Yes, Jev, OpenAI Decisions, open-weight models and Claude Haiku 5.5.
Related projects
Korean course code for attaching Jev to Claude Code, Codex and OpenClaw: skills, hooks, routers and open-model comparisons.
ProjectproductionJavaScriptClaude skill that gives a ten-line gut-check on whether Jev could help with what you are building, citing tested patterns.
Agent skillbetaMarkdownThirty worked Jev uses in ten levels, from a smart if statement to a pi coding agent that writes its own Jev questions.
ProjectbetaTypeScriptAgent skill that makes a coding agent try Jev first for bounded judgments like classify, rank or gate, and log each attempt.
Agent skillbetaPythonPublic Jev lab centred on a chess match broadcast where Jev picks from legal moves, plus smaller judgment booths and a Fast Flag catalogue.
ProjectprototypeJavaScript, Python25-second Japanese IT-buzzword game that shows Jev's speed and cost: every typed term is classified live, with no LLM-generated text.
ProjectprototypePython, JavaScript