stable-worldmodel is a PyTorch research library and CLI (swm) for world-model-based reinforcement learning. It provides an end-to-end research pipeline: parallel environment rollout and dataset collection (World + EnvPool), multi-backend episodic dataset storage (Lance default, plus HDF5, folder, video, LanceVideo, and LeRobot adapters), offline training of reference world models (LeWM, PLDM, PreJEPA, GCRL, TD-MPC2, Genie), and planning-based evaluation through model-predictive-control solvers (CEM, iCEM, MPPI, predictive sampling, gradient/PGD, Lagrangian). It is a hybrid of a structured software framework and research code: the package ships a CLI, CI workflow, and roughly 50 test files, while scripts/ contains Hydra+Lightning training pipelines, expert SAC/TD-MPC2 collectors, data-collection shards, and visualization notebooks. The core library is cleanly modular. stable_worldmodel/data uses a pluggable format registry to decouple dataset I/O from backends; stable_worldmodel/planning decouples solvers from cost functions via runtime-checkable protocols (Costable, Constrainable, Dynamics, Objective) composed by a ShootingCostEvaluator; stable_worldmodel/envs registers a broad gymnasium environment suite (DMControl, Gymnasium control, Fetch, OGBench, PushT, TwoRoom, Craftax, ALE, RoboCasa) that exposes domain-randomization variation spaces; stable_worldmodel/world orchestrates vectorized rollouts and evaluation; stable_worldmodel/wm holds model implementations and research losses (SIGReg, VCReg, PLDMLoss). Checkpoint utilities save/load with config JSON from local paths or HuggingFace repos, supporting reproducible training runs. Strengths: the format registry, protocol-driven planning composition, extended Gymnasium spaces with init_value/constrain_fn, and the ReplayBuffer's dual Dataset/Writer interface are genuinely reusable components. Tests pin the planner refactor to bitwise parity with the old cost API, and CI runs a substantial suite. Weaknesses: it is research-grade, not production-grade — lewm/module.py and pldm/module.py are near-duplicates, many TODO markers indicate incomplete features, a known Lance wide-column writer bug is pinned xfail unresolved, robocasa.py hardcodes a ~/robocasa asset path, and optional backend dependencies (torchcodec, lancedb, imageio, LeRobot) can silently degrade format support if unavailable.