Jev directory
CommunityJev Demos & Experiments
Muscle Memory
A software factory that cleans AI-written code by combining Jev (System One), Gemini (System 2), and learned Semgrep rules. It verifies cuts with public and hidden tests, learns rules from verified decisions, and logs every decision.
- Stars
- 1
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Jev settled 80 of 84 model decisions on its own, about 16x faster and 80x cheaper than Gemini.
- Reflex learned rules made 11 decisions at 0.05 ms and $0 each.
- Two rules were learned live: restating_comment and swallowed_exception.
- 0 unsafe cuts shipped; 1 bad cut was caught by tests and reverted.
- Hidden tests in target/holdout/ are never shown to any model.
Questions about this entry
Short answers drawn from the linked project or page.
Muscle Memory is a software factory that cleans AI-written code by combining Jev (System One), Gemini (System 2), and learned Semgrep rules. It finds sloppy code candidates, decides whether to cut them, verifies cuts with tests, and learns rules from verified decisions.
It finds candidates in code, then applies learned reflex rules, Jev atomic yes/no questions, and Gemini only when a Jev answer is unsettled (probability between 0.3 and 0.7). The keep/cut decision is a readable truth table in factory/questions.py.
Jev is the System 1 tier. It answers small, concrete questions such as whether an except block throws an error away. In the real run, Jev settled 80 of 84 model decisions on its own, about 16x faster and 80x cheaper than Gemini.
After 5 verified cuts of one kind, Gemini writes a regex rule. The rule must match at least 80% of verified cuts and zero verified keeps, or it is rejected. Rules are saved as Semgrep YAML in rules/ and committed to git.
The rule is switched off, and Gemini writes a v2 that excludes the problematic case. The digest shows mm-swallowed_exception-v1 was demoted after over-cutting a hunk the tests needed, then patched as v2.
It runs the code's public tests plus a second set of hidden tests in target/holdout/ that are never shown to any model. A cut ships only if public tests stay green and hidden-test failures do not increase.
Reflex learned rules made 11 decisions at 0.05 ms and $0 each. Jev made 80 decisions at 127 ms and $0.00004 each. Gemini 4 Hydrogen made 4 decisions at 2,031 ms and $0.003 each.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Vercel's fx safety reviewer, measured at 18× faster
Reddit: Early experiments wiring Jev into Pi as a safety layer and model router.
X: Decision outputs in Grep
Hacker News: jevmem keeps Claude Code project memory by asking Jev which turns are worth saving.
avinashjetwani61 pointsJevmem – automatic project memory for Claude Code, built on JevRead the thread on Hacker NewsMore Jev resources like this
Spanish-language prototype that uses Jev to classify event sign-ups and decide program, action, and message before sending, with a guardrail that approves, flags, or blocks outgoing text. Includes synthetic datasets, evaluation scripts, and a local engine that runs without a key.
Jev Demos & Experiments#madewithjev#demo#python
Communityblockstars-labs/30x-JEV
A Hugging Face Space by abidlabs that finetunes a Jev-style spam decision model and tracks the run with Trackio. The Space is hosted on Hugging Face and sleeps after inactivity.
Jev Demos & ExperimentsPlayground#demo#classification#madewithjev
Community
A Python demo that screens PubMed abstracts for one systematic review question using Jev. Six narrow questions per abstract go to Jev in a single call, with rule-based tiers and confidence gates deciding include, exclude, or human review.
Jev Demos & Experiments#demo#python#madewithjev
Communityanondo1969/jev-evidence-triage
Portuguese-language Google Colab notebook that routes delivery support tickets with Jev's Score and Choice primitives, turning report text, user history, and financial volume into a 1-10 complexity index and a deterministic lane choice.
Jev Demos & Experiments#madewithjev#python#routing
Communityjoaohppenha/otimizador_jev
Minecraft survival bot built on Mineflayer that asks Jev four typed questions every 600ms (threat level, action, whether to eat, flee direction) plus a tactical goal-selection loop, with simulated emotions and persistent memory feeding each request.
Jev Demos & Experiments#demo#games#madewithjev
Communityakash-kamat/jev-craft
An autonomous research agent that searches grant sources daily, scores each lead's fit with Jev, and emails one ranked HTML digest with deadlines, amounts, and apply links. Python runs the loop; Composio supplies the search, page-fetch, and Gmail tools.
Jev Agents & MCP#agents#madewithjev#python
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.