What is Médula and what problem does it solve?

It is a small, fully reproducible lab for measuring how coding agents coordinate on one repository, and a kernel that coordinates them: before every write it asks a fast decider whether the change collides with what the other agents are doing.

What did the first run matrix find?

Git merges the six branches without a single code conflict and the result is still broken: /export calls login(username, password) after login started requiring a second factor, with 6 red tests in every mode B run. With Médula both real conflicts are caught in every run with no unnecessary blocks.

How does Jev fit into Médula?

Jev is the fast decider the kernel consults before every write request. It decides in about 0.3 s for $0.00006, but 61% of its decisions on real write requests land in the not-sure band and escalate to the slow path, Sonnet and then Opus.

Can I try it without an API key?

Yes. The README's two-minute path runs entirely offline with a mocked provider: git clone the repo, uv sync the project and its demo-app and medula packages, run uv run pytest in medula, and run bench/self_check.sh.

What evidence is published?

Every run is published raw: agent sessions, diffs, evaluations, and each decision the kernel took. results/runs/d1/medula.db is a SQLite file with every question Médula asked and its answer, probability, latency, and cost.

What do I need to run it?

Python 3.12+ and uv as the package manager, with FastAPI, SQLite, and pytest for the kernel; Bash runs the bench scripts. The six tasks run on a room-booking API with Claude Code using anthropic/claude-sonnet-5 via OpenRouter.

What is the biggest open problem?

The not-sure band: 61% of Jev decisions on real write requests escalate to the slow path, which the README calls its biggest open problem. The repo lists concrete issues to claim, each with data behind it and a place in the code to start from.

Jev directory

CommunityJev Benchmarks & Evaluations

Médula

A reproducible lab measuring how several coding agents coordinate on one repository, plus Médula, a kernel that asks TypeSafe Jev before every write whether a change collides with what the other agents are doing.

See highlights Open github.com

Stars: 12
Language: Python
License: MIT
Reading time: 3 min
Published by: Community
Author: JoaquinRuiz
Last verified: Oct 2026

joaquinruiz/medula

User repository on GitHub

View on GitHub

Several coding agents, one repo, at the same time: a reproducible lab to measure how they coordinate, and Médula, a kernel that decides who waits for whom.

Forks: 7
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity catalog-import cobanov logicrw multi-agent coordination benchmark

Works forJev for Decision gating Jev for Verification

Used inJev for Software & DevTools

Languages: Python Shell
Platforms: CLI API SQLite
Capabilities: Open sourceRequires API key

Highlights

Branches hide semantic conflicts: git merges all six cleanly yet /export still calls login(username, password) after 2FA, 6 red tests in every mode B run.
With Médula both real conflicts are caught in every run with no unnecessary blocks (mode D) at the same total cost as classic per-file locks.
Jev decides in about 0.3 s for $0.00006, but 61% of its decisions on real write requests fall into the not-sure band and take the slow path.
Six tasks on a room-booking API, six coordination modes, the same agents (Claude Code on claude-sonnet-5) in every mode, raw runs published.
Offline tests need no key: uv sync, then uv run pytest runs 69 kernel tests with a simulated provider, a real server, and a hook script.

Quickstart

bash

git clone https://github.com/JoaquinRuiz/medula.git && cd medula
uv sync && (cd demo-app && uv sync) && (cd medula && uv sync)
(cd medula && uv run pytest)
bench/self_check.sh

Watch out

Copy reviewed against the repository README on 2026-10-15; it was truncated in the help-wanted section. The coordination matrix is small, the 61% not-sure rate is the README's own open problem, and tests need no API key only because the model provider is mocked.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

The trust ledger: measuring confidence your domain has earned

Oct 16, 2026

ASSAY-001 calibrated Jev on CLINC150 but found it overconfident on Banking77 — one audit, two verdicts. How calibration research measures the model-and-domain pair, discounts stated confidence, and signs the evidence.

Jev Benchmarks & Evaluations#calibration#evaluation#open-models

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai