Your own numbers show raw Jev argmax beating the mix 9/9 to 6/9 - why would I use the skill?

Because of what the raw hits were. One is ship_or_merge at confidence 0.29, and on a first timeout Jev's repeat_failure noul at 0.97 contradicts its own retry_same choice - the model disagrees with itself exactly where it matters. The mix is the conservative side of that disagreement, so it declines to act on the low-confidence call instead of merging something that should have waited. If you want the argmax, jev_raw is one of the three scored configurations.

Where do the priors and the question bank actually live so I can tune them?

In SKILL.md, not the README. The page only says SKILL.md covers when to call, the question bank, and the mix/policy, and the loop is driven by piping a state.json into scripts/loop.py. No prior values, thresholds or veto rules are reproduced on the README, so you would be reading the skill file to know what is being mixed.

Can I try this without a TypeSafe key?

Yes for the benchmark, no for the loop. python3 scripts/benchmark.py runs on committed Jev answers with no key, which is how the published scores were produced; adding --live refreshes those answers and needs TYPESAFE_API_KEY. The live loop, python3 scripts/loop.py < state.json, also needs the key exported plus the skill symlinked into ~/.grok/skills.

Is nine cases enough to act on?

Treat it as a fixture, not a benchmark. The README describes nine anonymized cases from one real session, plus a historical five-case set where the skill went from 0/5 to 4/5, and names the remaining misses as over-cautious gather_facts steps. There is no second fixture, no seed sweep and no confidence interval, so it shows what the mix changed on those cases rather than how often.

How do I install the skill?

Clone the repository, mkdir -p ~/.grok/skills, symlink the clone into ~/.grok/skills/jev-builder-loop, set TYPESAFE_API_KEY, then pipe a state.json into scripts/loop.py at a loop boundary.

What are the three configurations the benchmark scores?

Before is what a coding agent did without the skill, jev_raw trusts Jev's next_act argmax, and skill is the mix plus a hard-fact veto. Scores land in benchmarks/RESULTS.md.

What does the loop do each cycle?

You write the questions and the environment state, Jev returns typed probabilities, the script mixes them with your priors into posteriors, and then you act on the result rather than on raw model text.

CommunityJev Agents & MCP

Jev Builder Loop

Q: Is nine cases enough to act on?

Treat it as a fixture, not a benchmark. The README describes nine anonymized cases from one real session, plus a historical five-case set where the skill went from 0/5 to 4/5, and names the remaining misses as over-cautious gather_facts steps. There is no second fixture, no seed sweep and no confidence interval, so it shows what the mix changed on those cases rather than how often.

Q: How do I install the skill?

Clone the repository, mkdir -p ~/.grok/skills, symlink the clone into ~/.grok/skills/jev-builder-loop, set TYPESAFE_API_KEY, then pipe a state.json into scripts/loop.py at a loop boundary.

Q: What are the three configurations the benchmark scores?

Before is what a coding agent did without the skill, jev_raw trusts Jev's next_act argmax, and skill is the mix plus a hard-fact veto. Scores land in benchmarks/RESULTS.md.

Q: What does the loop do each cycle?

You write the questions and the environment state, Jev returns typed probabilities, the script mixes them with your priors into posteriors, and then you act on the result rather than on raw model text.

A Grok skill where you write the questions and environment state, Jev returns typed probabilities, and the script mixes them with your priors into posteriors before you act. Ships a nine-case benchmark you can run without a key.

See highlights

Open github.com

Stars: 3
Language: Python
License: MIT
Published by: Community
Author: rainbowpuffpuff
Last verified: Oct 2026

rainbowpuffpuff/jev-builder-loop

User repository on GitHub

View on GitHub

Grok skill: Jev as a judgment sensor in a builder-agent loop (priors × probabilities → next act)

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A Grok skill: you write the questions and environment state, Jev returns typed probabilities, you mix them with priors into posteriors, then you act. Jev is not a coder. The loop exists to cut wasted retries and operator babysitting. Set TYPESAFEAPIKEY. Then at a loop boundary: See SKILL.md for when to call, the question bank, and the mix/policy. Problem: ship a signed Android App Bundle after a UI change merged to main.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Jev is not a coder: you write the questions, it answers them, and SKILL.md mixes probabilities with priors.
scripts/benchmark.py replays committed Jev answers with no API key; --live refreshes them.
Three configurations scored on one fixture: before, jev_raw (next_act argmax) and skill (mix plus veto).
Nine anonymized cases from a real session: a green PR waiting for review, Gradle retried with empty JAVA_HOME.

Watch out

Read the benchmark before trusting it. The same README reports skill at 6/9 and jev_raw at 9/9 on the nine-case fixture, so the mix scores lower than raw argmax here, and one of the raw hits is ship_or_merge at confidence 0.29 - and on a first timeout Jev's repeat_failure noul at 0.97 contradicts its own retry_same choice. The skill sits on the conservative side of that disagreement, which is a defensible trade but a trade. The nine cases come from one anonymized session, and the only other figure given is a historical five-case set going 0/5 to 4/5. The priors, the question bank and the hard-fact veto rules are not in the README - they are in SKILL.md, so tuning happens outside what this page shows. Running the loop needs TYPESAFE_API_KEY and a Grok skills directory layout with a symlink into ~/.grok/skills. No token counts, per-decision latency or pricing are published, and there is no second fixture or seed to check whether the scores hold. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-recruiter

Local LinkedIn recruiting workspace on Browser Use's Jev Ultrafast: Jev screens titles, drives navigation, and assesses profile excerpts against your brief while a live browser stays visible; matches and evidence export as JSON.