H3 HyperFlow · TAIL Effects Execution Audit (T8 EXP)

Most ComfyUI "audit" nodes are decorations. This one isn't. It reads the immutable effect coverage recorded inside a completed TAIL result - the actual Relay and EAV calls that happened during sampling - and reports them, passing the typed result straight through so it costs you nothing structurally.

Why this exists: an Apply node in a graph proves only that you wired something up. It might have run before you changed the config, it might be bypassed, it might be a leftover counter from a cancelled run. The pack's answer across every route is to record effect calls inside the completed result at sample time, and have the audit read that record instead of any node's mutable state. Same idea as the P7 stage receipts: the evidence travels with the output.

Where it fits

This is the continuous HEAD/TAIL route, not P7. In that route the sampling is split at an absolute interval - HEAD runs 0:split capturing a raw model-space AV x_sigma, TAIL runs split:8 injecting that state with no new noise and no resize - and both halves can carry their own external effects. The TAIL audit is the one you place after the TAIL sampler, before storage or decoding.

TAIL Only sampler → TAIL Effects Execution Audit
   ├→ completed_result → Save / Load
   └→ av_latent → VAE decode → delivery

Inputs and outputs

One input: completed_result, the typed T8_HYPERFLOW_COMPLETED_AV. If you're feeding a plain LATENT in here you've skipped a step - the typed result is produced by the TAIL sampling path (and by Load Exact Completed TAIL AV if you're reading from disk).

Outputs: completed_result passes through unchanged so you can chain storage after the audit, av_latent gives you the final AV for decoding, and report_json is the interesting part - the coverage report.

What the report is not

It is not a quality verdict. The node says so, and it's an honest accounting of what it can know: "these effects executed over this coverage" is a fact about the run; "this segment looks right" isn't. The pack's whole experimental section is careful about separating machine checks from human review, and this is one of the places where it's written into the node description.

Install

Manager → search 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 of ComfyUI and a browser refresh - this pack registers a large node set and partial restarts are the usual reason a node "doesn't exist". requirements.txt deliberately installs no packages so the install can't interfere with ComfyUI's Torch/CUDA. For the continuous route you need H3 base weights in models/diffusion_models, the Qwen text encoder in models/text_encoders, VAEs in models/vae, and the original non-pruned HyperFlow adapter in models/hyperflow/loras/ - loaded via the pack's dedicated HyperFlow loader, not a generic LoRA loader.

How to use it productively

Two habits. First, put it on the path you'll actually deliver from, so a completed run can't reach storage without its coverage being recorded. Second, read it when something looks different from your last run: if you enabled an EAV gain and the output looks identical, the report is the fastest way to find out whether the effect really applied or you'd left the config at the report-only default. That default exists precisely so that "apply" is something you choose, not something you get.

Beyond effects, this node is also just a convenient typed waypoint - the pass-through means you can hang the audit in the chain without rewiring anything, and the av_latent output lets you decode the final AV without going back to the sampler.

One caveat about the venue: the continuous route is genuinely experimental in this pack, and the docs say the support matrix is still filling in - more backends, full assets, browser save/reload and human quality review are all explicitly listed as pending. Treat a green run as "the mechanics work", which is what it is.