matlab-agentic-ml

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.

MATLAB ML/RL Workflow

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.

Public video learning catalogs

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.

Workflow

1. Pin MATLAB release, toolbox availability, session owner, and random seed.
2. Define the data split or RL environment before training.
3. Train with recorded options and elapsed-time stages.
4. Evaluate on data/episodes disjoint from training and record the evaluation metric separately from the training metric.
5. Store a result artifact with schema, units, release, seed, split/environment, digest, and one to four dominant timing stages.
6. Replay the artifact through 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.