Track
muxnode.ai
Plane
Media and essence
Charter
v0.4
Status
Draft, pre-validation

A CueSeal track

MuxNode: what quality costs

Research on encoding, packaging, and delivery, for engineers running their own pipeline. Part of CueSeal.

01

The problem with most encoding advice

Most encoding advice is a preset somebody got lucky with once. Proper measurement is tedious, and the usual tools will produce a confident wrong number without a warning.

This track does the tedious part: what a given quality level costs in bits, for content that is not a broadcast test clip. What self-hosted delivery costs next to handing the stream to a platform. Which measurement habits produce a wrong answer quietly, and what catches each one. And whether provenance and metadata survive a transcode.

02

Open questions

  • What does an encoding ladder cost, in bits, at an acceptable quality floor, for footage that looks nothing like a broadcast test sequence? Screen capture, handheld video, and webcam streams behave differently from the clips most comparisons use.
  • What does self-hosted delivery cost in latency and bandwidth, next to letting a platform handle it? A measured number settles whether owning the pipeline is realistic.
  • How many ways can a standard measurement produce a wrong answer without looking wrong? A silent pixel format change, a colour range shift, a scaler nobody noticed: each one returns a clean number that means nothing.
  • Does a C2PA manifest or a caption track survive a normal transcode and re-wrap, and when it does not, which step strips it?
  • Does a model reading per-frame bitrate residuals directly find anything the metadata-level predictor misses?

03

About the tools

FFmpeg does most of the work here, with the open encoders and quality metrics around it, VMAF included. It will measure the wrong thing without complaint if allowed to. Half the value of this track is a set of checks built to catch that before a number reaches a publish step.

Licensing is stated plainly, since it changes what can be built on top of a result.

04

Right now

Where this actually stands

First piece of work in progress: a set of reproducible cases where a common measurement produces a confident wrong answer, plus the check that catches each one. Shell scripts and FFmpeg, runnable against most pipelines.