H3 HyperFlow · HEAD Effects Execution Audit (T8 EXP)

A one-input, two-output node whose entire job is to read a number you can't get anywhere else. It takes a continuous_boundary from a HEAD stage, reads the immutable effect coverage recorded during that stage's real sampling, and passes the boundary through untouched.

If you've bound an EAV effect to the HEAD half and want to know whether it executed, this is the only trustworthy answer in the graph. The Apply node's counter tells you what you asked for. This tells you what the model did.

Why the difference is real

The continuous HyperFlow route attaches to a model that's already been patched by the HyperFlow loader, and possibly by content LoRAs, a Relay conditioning wrapper and an attention backend. The pack's stated policy is to preserve and delegate unknown patches rather than reject them - which is the pragmatic choice for an ecosystem where everyone stacks optimisations, and which also means an effect can be quietly bypassed by a wrapper that reached the same hook first.

So the audit reports coverage from the completed stage's record. Missing or bypassed effects come back as not-verified rather than as a pass, and the author's line is blunt: connection proves no effect. A report that says unverified is doing its job, not failing.

It doesn't certify quality, either. Complete coverage with an effect that makes your clip worse is a perfectly normal outcome, and the audit has no opinion about it.

Inputs, outputs, and why pass-through matters

In: continuous_boundary. Out: continuous_boundary (the same object, so the chain continues) and report_json.

The pass-through design means you can sit this node between the HEAD stage and whatever consumes the boundary - a save node, the TAIL stage, another audit - without rewiring anything. That's deliberate; an audit you have to restructure your graph to use is an audit nobody runs.

Put it after the HEAD Only sampler, and before the HEAD stage gets saved. If you save a HEAD artifact that had effects bound, the effects' identity travels with it; on resume you only execute the TAIL half, and the docs are clear that the effect tag on a loaded artifact describes where it came from rather than something being re-applied.

Reading the report

Look for actual calls, not intent: the audit distinguishes real routed attention and forward coverage from an empty record. If you're comparing two runs to decide whether an effect earns its keep, this plus a fixed seed is the honest way to do it, because the alternative - squinting at two clips and deciding one looks more "cinematic" - is how people end up shipping placebo settings for months.

Install

Manager → MiniMax H3 Audio T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Full restart, recent ComfyUI core, H3 base model, Qwen3-VL encoder, both VAEs, and the original HyperFlow adapter in models/hyperflow/loras/. No auto-download for the adapter, no pruned fallback. The pack's requirements file is intentionally empty - nothing here will touch your Torch/CUDA install.

Where people get burned

Auditing the wrong half. This reads HEAD coverage; the TAIL half has its own audit node, and the two aren't interchangeable. If you've configured different effects per half - which you can - you need to look at each.

Bypassed effects read as not verified, and people sometimes read that as an error and go hunting for a bug in the audit. It isn't a bug. It means a wrapper owns that entry point and your effect didn't get a call. The honest options are to change the stack, or to accept that the effect is decorative in this configuration.

And keep CFG at 1 when effects are active. If you raised CFG without supplying a true negative conditioning, you're outside the qualified configuration and the coverage report is the least of your problems.