H3 Context Validator

Every prompt tool that silently "cleans up" your text is a tool you can't trust, because you can never tell what it changed. H3 Context Validator does the opposite: two typed inputs, a diagnosis, and not one character of your prompt touched. Its description says it plainly - it audits prompt structure and hard constraints "without auto-correction."

That's the whole personality of the node, and it's why it's the last stop before generation in the README's minimum route: Request → Plan → Compiler → Validator → Preview.

What goes in

Both inputs are required and both are typed objects, so there's no way to hand it a string and hope:

• plan (H3_CONTEXT_PLAN) - the typed plan under audit, from H3 Context Plan (or the accepted plan out of the constrained-semantic chain).
• prompt_document (H3_PROMPT_DOCUMENT) - the rendered document. This comes off the third output of either H3 Context Compiler or H3 Source-Profiled Renderer. New users wire the prompt string here and get invalid_prompt_document; the string is output 0, the document is output 2.

Hand it anything else and it fails closed with invalid_plan or invalid_prompt_document. It also defers quietly during ComfyUI's pre-execution validation pass, when linked inputs are still unresolved - that's normal, not a bug you need to report.

What happens inside

It builds a draft report from the plan/document pair and pushes it through the report lifecycle: draft, validated, with the diagnostics attached, and the result registered against the same reconstruction chain the rest of the pipeline uses. If any of that breaks, you get validation_failed and nothing else - no partial verdict, no "probably fine."

Severities are graded, which is the part worth paying attention to. Errors are things you must fix; warnings are uncertainties you may accept, and the README frames the difference exactly that way. A timecode from the plan that got dropped in rendering, a reference label in the document that doesn't match the registry's order, a hard constraint that no longer appears - those are structural claims the validator can actually check, because it's comparing two typed objects rather than reading prose.

What it can't tell you: whether the video will be any good. The pack is unusually careful about this. Validation grades the paperwork. Separately, "guide readiness" asks whether the plan covers what the official H3 prompt guide expects - and Incomplete there is a legitimate state for a prompt that's structurally valid. Neither verdict is a quality score, and treating a green validation as proof your clip will be good is the fastest way to be disappointed by a video generator.

The two outputs

• validation (H3_VALIDATION_RESULT) - the verdict itself, with its diagnostics.
• validated_report (H3_CONTEXT_REPORT) - the report other nodes want. Wire it into H3 Context Preview to read it, into H3 Native MiniMax H3 Adapter to hand the prompt to the native H3 nodes, and into H3 Local Reconstruction if you're building the full acceptance route.

# the bundled working example, if you'd rather read a graph than a doc
# workflows/m7_04_h3_context_provider_transparency.json
#   1 Request -> 2 Plan -> 3 Compiler -> 5 Validator(plan=[2,0], prompt_document=[3,2])
#   -> 7 Native H3 Adapter -> MiniMaxH3ImageToVideo

Where people get burned

Editing the prompt text after validating and then queueing anyway. Now your validated report describes a prompt that no longer exists, and the pack has a node for exactly this situation: H3 Context Audit Override records a deliberate decision to proceed with a prompt edit while keeping the original diagnostics visible. Use that, and the disagreement is on the record. Free-hand edit the graph and it isn't.

Second, don't expect the Validator to be the place a missing first frame shows up. Mode-versus-assets complaints belong to the plan/normalization path - the Validator audits the plan/document pair it was given.

Install it

Not in the Comfy Registry yet as of the README, so clone:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-MiniMaxH3-Studio.git

Restart ComfyUI, then Add Node → h3_context → validation. The pack ships no runtime Python dependencies, downloads no models, and building plus validating prompts needs no H3 weights - so you can run the whole prompt half of a workflow on hardware that can't generate a single frame. Same author as the ComfyUI-Doctor diagnostics pack, which is arguably why the error text here is worth reading rather than skimming.