H16-3 · Window EAV Actual Calls (T8 EXP)

Here's an uncomfortable thing about ComfyUI effect nodes: most of them report what they were asked to do, not what the model actually executed. You wire up an attention-routing effect, the node UI says "applied", the picture looks subtly different, and you have no evidence any of it happened - because a wrapper further down the chain may have taken the same hook and skipped yours.

This node is the receipt. It reads the immutable call records captured during one completed H16 PASS2 window and tells you how many EAV forward passes and routed attention calls actually ran after that window.

What "EAV" means here

In this pack, EAV is the stage-local effect route - the external attention/effect wrapper the author builds on top of the existing Enhance-A-Video/FETA machinery. You configure it with the pack's stage EAV config node, bind it to one exact window with MiniMaxH3H16EAVApplyEXPT8 (which outputs a runtime handle), and then sample that window. The window node itself doesn't tell you whether the effect survived contact with the model. This does.

Note the deliberate asymmetry: the Apply node hands you a runtime object, and the audit consumes that exact runtime plus the window result and segment spec. So the audit is checking this run's numbers against this window's identity - not reading a global counter that some other graph has been incrementing.

Inputs and outputs

Three inputs, all typed, none of them a widget:

• window_result - the typed result from the H16 window you want audited.
• segment_spec - the window identity (index, plan SHA, source identity) so the audit can confirm you're auditing the window you think you are.
• runtime - the runtime owned by the EAV Apply node for that same window.

Out: cumulative_av_latent and report_json. This node is an output node, so it's meant to be a terminal in the graph, and the report also surfaces in the node's text UI. The report is where the numbers live - actual forwards, actual routed attention calls, and whether anything came back as uncovered.

One detail worth knowing: the cumulative_av_latent it hands you is the H16-selected final audio, not the core video's first-pass-audio intermediate. If you're comparing audio between runs, compare like for like.

What it can't tell you

Almost every comment in this family of nodes boils down to one sentence: connection alone proves no effect. Wiring the audit node to an EAV Apply node proves you wired it. The report showing unverified or zero calls proves the effect did not get used - that's the useful direction. Unknown third-party wrappers and backends still execute and get archived, but they won't be certified for reuse, and the author won't silently substitute a dense path to make the numbers look pretty.

It also doesn't certify quality. A window can report perfect coverage and still look worse than the one without effects. This is an instrumentation node; treat it as a scope, not a verdict. And if you're running Relay and EAV on the same window, use the joint EAV audit rather than the standalone Relay audit - the pack splits them and the Relay-only audit will not account for the EAV calls.

Install

Nothing special, it's part of the same pack. 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 afterwards, and keep core + frontend + Manager in step or the whole category goes red. The base pack installs zero extra packages by design; optional EXP features check their own dependencies only when used.

Where people get burned

Auditing the wrong window is the classic. segment_spec and window_result must come from the same window, and the runtime must be the one from the EAV Apply node that fed that window's sampler. Mixing window 3's result with window 4's runtime produces a mismatch error, which is the polite outcome.

The impolite outcome is the one this node exists to prevent: an effect silently bypassed by an unknown patch, a report you skim, and a 20-minute render you can't reproduce. Read the report on the first window before you commit to seven of them.