H3 Chunked v5 · Window Relay Actual Calls (T8 EXP)

Attention-routing effects are exactly the kind of thing that looks like it worked. The clip changes a little, you nod, you move on. Then someone asks you whether the routed attention actually executed on window 1, and you have no evidence at all.

H3 Chunked v5 · Window Relay Actual Calls is the receipt. It audits actual routed Relay attention calls after one matching v5 PASS2 window and writes it to report_json. Output node, no downstream consumers, run it and read it.

Wiring

Three required inputs: window_result (the typed result from the PASS2 window you're auditing), prepared (the same shared global preparation that window consumed), and runtime - the T8_CHUNKED_V5_RELAY_RUNTIME returned by H3 Chunked v5 · Project External Relay to Window for that window.

Outputs: cumulative_av_latent and report_json.

The runtime object is bound to a specific source, lift, plan, window index and binding hash. That's what lets the audit tell you it audited this window and not its neighbour. Swap windows without rebuilding the projection and the audit refuses rather than reporting the wrong counters - annoying in the moment, and much better than believing numbers that belong to a different render.

What it does not certify

The description draws the line itself: this standalone audit does not certify Relay+EAV composition. If both an EAV apply and a Relay projection were bound to the same model, the combined calls belong to the EAV audit's story, not this one. The pack refuses to add two half-claims and call it a whole one. If you're combining effects, be aware you've left the documented path.

It also won't tell you anything about quality. Call counts are coverage evidence, not a verdict on timing, lip-sync or whether your dialogue event landed on the frame you meant. Relay events are re-gridded in absolute motion time, which is the mechanism for that correctness - the audit only confirms the routing ran.

Practical advice the pack itself gives: change one variable at a time. Same seed, same recipe, report-only versus applied, then compare. Otherwise you're looking at a different clip and calling the difference an effect.

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

Fully restart ComfyUI, then refresh the page - this family's nodes need a real restart to appear, and a queue refresh does nothing. Recent Core with native H3 support is required; Registry and GitHub release cadence are independent, so clone from GitHub if the Manager listing is stale. Base and EXP nodes add no pip dependencies by design: the pack's requirements.txt is empty so it can't replace ComfyUI's Torch/CUDA. Weights (main model, Qwen text encoder, video VAE, audio VAE, plus the learned 3D upscaler from the author's t8star HuggingFace repos) are separate downloads.

Where it goes wrong

Counters that don't move. Usually the Relay model wasn't the one you wired into the window - pass the model and positive outputs of the projection straight through, not the raw model.

Refusals after an edit. Editing the plan, the window index or the Relay text after building the projection invalidates the runtime binding. Rebuild forward.

Assuming report_only EAV is the same thing. It isn't; EAV and Relay are separate effect families with separate audits on this route.

Nodes missing entirely. Check custom_nodes for a second, stale copy of the pack. Old test copies serve old schemas and shadow the real install. Rename the stale directory with a .disabled suffix - a leading underscore does not disable a node pack - then restart and confirm the python_module in /object_info.