Classify every source before loading the graph

Keep three evidence layers separate. The sanitized the public discussion signal shows that the phrase circulated during the recorded window. MiniMax's repository snapshot checked on August 15, 2026 supplies the limited first-party record for H3 model variants, checkpoint and hosted-module boundaries, and license links. A downloaded ComfyUI graph, custom-node bundle, install note, or demonstration remains a third-party community artifact unless its own publisher and lineage say otherwise. Save the graph URL, publisher, retrieval date, file hash, stated license, revision identifier, and accompanying documentation before opening it. Do not infer that MiniMax or SEELE authored, endorsed, integrated, secured, or tested the graph merely because it names H3.

• the public discussion signal: bounded circulation only
• MiniMax repository: dated H3 facts only
• Community graph: separately attributed artifact
• Unknown dependencies stay unknown until resolved

Turn the canvas into a dependency manifest

Read the serialized graph and list every node type, extension repository, commit or release, model file, encoder, decoder, audio component, preprocessing step, path assumption, environment variable, and external service call. Record the ComfyUI build plus the Python, framework, accelerator, operating-system, and memory context required by the publisher when those details are available. Mark absent versions and unexplained files as blockers instead of choosing convenient substitutes. Review custom-node source and install scripts before execution, isolate the environment, use least-privilege credentials, and never copy cookies, tokens, proxy settings, or private user data into a public workflow record. An attractive canvas is not a reproducible specification until each executable dependency has an accountable source and a compatible version range.

Run a clean-room reproduction with a bounded conclusion

Use an authorized fixture that does not contain private or unlicensed material. Rebuild from the manifest in a fresh environment, capture dependency resolution, graph validation, model loading, inputs, visible settings, random-state behavior when exposed, console output, failures, generated files, and reviewer notes. Compare the reproduced graph hash and resolved-node inventory with the reviewed artifact before judging any result. A useful disposition is reproducible, conditionally reproducible, blocked, or not reproducible, followed by the exact reason and retest trigger. One successful run does not prove quality, safety, repeatability, general compatibility, current hosted availability, or SEELE support; it only documents what occurred in that dated environment.