dsh-wm

In DeepSeek Harness, researchers often struggle with lacking intuitive tools to visualize and evaluate world model outputs effectively. DSH-WM, a playable toolkit, addresses this gap by enabling users to analyze rollout frames, identify model routes, score runs, and iterate on the research loop without requiring a GPU or complex setup, making it accessible for quick starts. Core capabilities include frame inspection via wm_inspect, comparison pages with wm_view, three world model route knowledge (3D display, pixel video generation, latent prediction), run scoring, and an RSI-based iteration cycle, where users can visually compare predicted vs. ground-truth frames, use knowledge cards for guidance, and refine research processes with skills like wm-rsi. This plugin is ideal for world model researchers, AI developers, and DeepSeek Harness users who need efficient model evaluation, reproducible experiments, and streamlined research workflows, simplifying assessment and iteration steps to accelerate innovation in the field.