MiniMax H3 RAVEN Request Audit / 流式请求审计 (Advanced EXP/T8)

RAVEN Streaming samples H3 causally, chunk by chunk, so it can reach frame counts ordinary sampling can't touch. But its supported surface is narrow - strictly T2VA (text-to-video-audio), strictly an empty AV latent, strictly an unpatched causal model. This node is the contract checker that sits between your graph and the external RAVEN sampler and enforces that narrowness. It passes the model, conditioning, and latent through only after the checks, and hands you a compatible boolean and a decision so nothing sneaks through silently.

It's the middle of a three-node T8 RAVEN chain: RavenStreamingProfile emits the parameters, RavenGuardedLoader preflights the model and memory, and this Request Audit validates the request itself. The mechanism: it delegates the strict conditioning, empty-AV-latent, and causal-MODEL checks to the external runtime, verifies the mandatory 266-module adapter is present, and blocks unreviewed model/attention patches. Mechanical conflicts are things like references or keyframes in the conditioning, a non-empty latent, masks, a non-causal model, a missing adapter, or external object patches.

Inputs that matter

• model, positive, latent - the trio going into the sampler. Required.
• steps (4), video_shift (12), audio_shift (3) - the published defaults, shifts tuned for the 4-NFE flow trajectory.
• sink (2), window (2), kv_cache_storage (cpu_pinned) - RAVEN's chunking and KV placement. cpu_pinned is the reviewed default; gpu KV is your own adventure.
• allow_experimental_over_192 - RAVEN's grid reaches 362 frames, but anything above 192 is explicitly experimental and demands this acknowledgment flag. Default false.
• enforcement - block_outside_reviewed_envelope by default; the tooltip lists the mechanical conflicts it blocks on.

Outputs: model, positive, latent (passed through), plus compatible (bool), decision, and report_json.

Installing

Part of the T8mars MiniMax H3 Audio T8 pack (Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes), and it needs the separately installed RAVEN streaming plugin alongside it. No pip deps from this pack. Full BF16 H3 base + RAVEN LoRA required, same as the Guarded Loader.

The take

If you're assembling a RAVEN stream from scratch, this is where you'll spend your debugging time. The common failure is wanting I2VA or Ref2VA behavior - a reference image or a non-empty latent - and getting a clean block because RAVEN v0.1 only audited T2VA. That's not the node being stupid; that's the node honoring the fact that only the T2VA causal path has been reviewed. decision and report_json will tell you exactly which check failed, so read them instead of flailing. And keep allow_experimental_over_192 off until you've actually run inside the reviewed range. On the author's own machine this whole route refuses to run, so treat any pass as "it might survive", not "it will be good."