Automated Video QC

Automated video quality control (QC) is the practice of inspecting a finished video file with software before it is delivered, to catch technical defects such as black or frozen frames, loudness and peak violations, missing audio, caption problems and format errors. Broadcasters and streaming services have required file-based QC for years, because a defect found after delivery is expensive. In AI video editing and generation pipelines, QC takes on a second job: checking the rendered file against the plan the system meant to make, instead of trusting an agent's report that the job succeeded.

What Broadcast QC Checks

Professional delivery runs on shared definitions. The European Broadcasting Union's QC work, first published as EBU Tech 3363 and now maintained online at QC.EBU.IO, catalogues hundreds of test definitions, each with references, tolerances and an example, so that different QC tools give comparable results. In the UK, delivery to broadcasters under the DPP's AS-11 specification requires a photosensitive-epilepsy (PSE) test with an approved device such as Harding FPA, and Ofcom's Broadcasting Code requires broadcasters to keep the risk to viewers with photosensitive epilepsy low. Streaming services publish their own delivery specifications; Netflix open-sourced Photon, a validator for the IMF packages it receives.

Commercial file-based QC platforms such as Telestream's Vidchecker and Qualify, Interra Systems' BATON and Venera's Pulsar automate these checks at scale, with templates for broadcaster and streamer specifications, PSE tests and, increasingly, machine-learning checks.

The checks fall into a few families:

• Picture: black frames, frozen frames, flash frames, out-of-range levels or colors, interlacing errors, wrong resolution, frame rate or aspect ratio.
• Sound: integrated loudness and true peak against a target (see loudness normalization), silence where there should be sound, clipping, wrong channel layout, phase problems.
• Text and graphics: captions present and in sync, readable and not too fast; titles and graphics inside the platform's safe area.
• Conformance: the codec, container, duration, frame count and metadata the delivery specification requires.
• Safety: in some markets, flashing and patterns that could trigger photosensitive epilepsy.

The Open-Source Toolkit

Most individual checks are available in open-source tools. FFmpeg includes filters that detect black intervals (blackdetect), frozen video (freezedetect), silence (silencedetect) and interlacing (idet), measure loudness to EBU R 128 (ebur128), and report per-frame picture statistics, including pixels outside the broadcast range (signalstats). Its companion ffprobe reports the container, codec, frame rate, duration and frame count. A single decoding pass with several of these filters is enough for a useful first-line QC report. What open-source tools don't supply is the judgment layer: which findings matter for this delivery, and which ones were intended.

Why AI Pipelines Need QC

An AI agent that edits or generates video will report that it finished. That report is text the model produced, not a measurement. Research on agents shows why this matters: a 2026 study of agent failures found that "false success", an agent confidently reporting a task done when it wasn't, made up 44 to 52 percent of failures on one benchmark and 75.8 percent on another, and that no model used as a judge reliably caught it. A render job that exits cleanly proves only that a file was written. The video may still drop a line of dialogue, repeat a slice of audio, hold a frozen frame where a shot failed to load, place captions over a speaker's face, or come out far too quiet. These are exactly the defects that are hard to see from the plan and easy to see in the file, and they matter more as generated shots enter the edit.

QC in an AI pipeline therefore checks the output against intent. Because the plan is structured data, as in video as code, the checker knows what the file should contain: how many frames, which spans are planned fades and which black frames are defects, where speech should be heard, where each caption and graphic sits. One AI video harness checks the rendered file against its plan: it decodes the real render, transcribes its audio locally to confirm that the planned speech is all there and in order, and labels each finding as observed (measured on the decoded file) or heuristic (inferred from the plan's geometry and conventions). Its QC never edits; it proposes fixes that go back through the same review and confirmation as any other change.

This is the media version of a broader rule in agent reliability: verify the artifact, not the agent's account of it. It is also where agent evals and production QC meet, since the same checks that gate delivery can score a pipeline over many runs.

Measured Versus Judged

Some QC is measurement: a frame count, a loudness reading, a span of digital black. Some is judgment: whether a caption is legible over a busy background, whether a crop has cut off a face, whether a dark shot is a defect or an artistic choice. Good QC reports keep the two apart and say which is which. Measured findings can block delivery automatically; judged findings are better surfaced with the frames that show them, for a person or a multimodal vision model to look at. Using a model to read sampled frames is a form of LLM-as-a-judge, and has the same caveat: a model's verdict is evidence, not ground truth.

Limits

• Intent is hard to infer. A detector can find a frozen frame but can't always tell a deliberate freeze from a stalled render; source evidence and the plan help.
• Conventions aren't guarantees. Platform loudness targets are often community-measured rather than published, and platforms normalize playback themselves.
• Speech checks depend on transcription. Names, music under speech and overlapping voices produce false alarms, so single-word slips are usually ignored.
• QC can't judge the story. It catches technical failures, not a weak cut. Editorial review still needs a person.