القائمة
Use the official MATLAB MCP execution layer for machine-learning and reinforcement-learning workflows, with explicit split, evaluation, release, seed, environment, and result-artifact gates.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Four-machine Radia release gate. Use when the user asks for release-qud, release_quad, release_qud, post-release deploy, GitHub Release publication, or the Definition Of Done for Radia releases. Coordinates PyPI and Simulink release candidates across LAB, 100号機, mdx, and hibino via tools/release_qud.py.
Publish Radia's Simulink application library after verifying its five application blocks and tracked MATLAB-only samples. Use simulink-app-health before deploy/release-qud; Cubit's toolbar remains a separate deployment surface.
Verify Radia's production Simulink application blocks for EM, PCB, Motor, Stream Function, and IH. Use after changing buildLibrary, applicationSFunction, runApplication, application.py, a DesignSpec/calc script, the application interface manifest, or the tracked Simulink library; also use when a block does not run or leaves incomplete artifacts.
Take inventory of the radia package's proprietary Python and pybind API surface, classify API families under the Reduce Proprietary API Surface policy, and produce a committed audit plus phased reduction plan. Use after public API changes, before deprecation work, or for periodic read-only surface audits.
Historical deploy notes. Current Radia deploy/release uses release-qud plus cubit-plugin-install; do not follow old .ccl/Cubit-2025.3 recipes except for archaeology.
Verify GMSH .geo + .msh files generated by Radia panels BEFORE opening in the GUI. Catches invalid options (Mesh.Volumes), duplicate element IDs, missing Merge targets, node count mismatches. Run AFTER any calc_*.py edit that touches GMSH export, BEFORE clicking Open GMSH.
| name | matlab-agentic-ml |
| description | Use the official MATLAB MCP execution layer for machine-learning and reinforcement-learning workflows, with explicit split, evaluation, release, seed, environment, and result-artifact gates. |
Use the official MATLAB MCP server as the execution substrate. This package adds small, solver-neutral artifact gates; it is not a second MATLAB runtime.
For AIcia Solid Project, inventory all public channel tabs rather than only the Videos tab. The 2026-07-20 snapshot contains 397 distinct IDs: 245 videos, two Shorts, and 150 streams. Public titles and metadata may nominate a CAE lesson, but transcripts and media are not copied. A candidate is not numerical ground truth and cannot be promoted without deterministic seeds, units/schema, provenance, and an independent forward-solver verification.
validate_matlab_ml_rl_artifact and, for the
v2 split/evaluation contract, validate_matlab_ml_rl_v44_identity before
treating it as a learning result.For the v46 generalization contract, also bind nonfinite filtering, worker
seed, restart state, RL episode timeout, environment-defined termination,
training-only exploration, checkpoint lineage, structured tool-argument shape,
existing-session attachment, timeout/error class, and monotone checkpoint
order. Treat any mismatch as needs_attention even when the numeric output
looks plausible.
For supervised learning, cross-validation or a holdout is required; a resubstitution score is not a validation result. For reinforcement learning, the environment, training episode count, evaluation episode count, and mean evaluation return are required. Evaluation must use the pinned environment, a fresh evaluation seed, and no exploration. A training curve alone is insufficient.
Use MathWorks' official MATLAB MCP/Agentic Toolkit for tool discovery, execution, testing, and code analysis. Keep public knowledge generic: do not embed lab paths, licensed solver results, or private MATLAB MCP implementation. The legacy Gypsilab acoustic lane is replay-only; new MATLAB learning belongs to this official ML/RL lane.