Jev Tune

The documented first command is jev-tune discover --model MyMotor.slx --json, which finds the gains, measured signal, and reference by following the signal graph. The tuning run is jev-tune matlab --model MyMotor.slx --auto --plot before_after.png --write-back. The package targets Python 3.10 and 3.12, with MATLAB R2024b verified.

No. It borrows the Jev interface, the three typed primitives choice, score, and noul with calibrated probabilities, but not the hosted Jev model or Laya's 421 M text encoder. The runtime policy is a 200 KB numeric head trained inside this repository; Laya is mentioned only as a text baseline that was never wired in.

jev-tune doctor --model explains in nine lines why a model will not tune: whether the file loads, whether it compiles, whether it uses referenced or protected models, whether Stateflow is present, which solver is used, and whether a loop can be found at all. Each check reports ok or FAIL with a short detail line.

No. The repository states it is not affiliated with, endorsed by, or sponsored by TypeSafe AI. It is described as Jev-inspired and uses the Jev interface only; the mapping of what is borrowed and what is not is documented in docs/JEV_MAPPING.md.

Independent loops, such as four axes or three zones, are tuned one after another with the --all-loops flag. A cascade is refused on purpose: the command returns cascade_suspect together with the loops it found, because tuning one loop of a cascade and reporting success would be worse than saying so.

Yes. AGENTS.md tells an agent to run jev-tune discover first, then the matlab command with --auto, --plot, and --write-back. When the model is ambiguous the command reports confidence: medium instead of guessing, and the agent asks one question. Claude Code, Codex, OpenCode, and Cursor are named as usable agents.

The documentation has a section stating that if you already have an accurate model, you should use pidtune. jev-tune is aimed at cases where the model is not fully known and the tuning budget is measured, with explicit abstention and a hard hold policy when the decision is uncertain.