A transcoded MP4 arrives from a creator, passes through storage, gets resized for social, then lands in an automation flow that pushes it into a CMS. By the time someone inspects the file, the simple tags are only part of the story. The operational questions are harder: which stream is the source of truth, did rotation survive remuxing, did timecode shift, and can you still prove where the asset came from?
The JSON stub below shows the problem. It has fields for container, streams, timecodes, validation, and provenance. If extraction stops at container.tags, the pipeline misses the data that breaks downstream jobs. I care more about stream-level facts such as codec, frame rate, pixel format, sample rate, color data, and time base, because those values decide whether a render farm accepts the file, whether a player syncs audio correctly, and whether an editor sees the expected timeline.
A practical extractor reads metadata in layers. Start with the container and stream map from ffprobe. Then inspect embedded timecode tracks, side data, and any transformation history your platform records. If provenance matters, read C2PA manifests too. A file can look valid in FFprobe and still fail a compliance check because the provenance chain is missing or broken.
ffprobe -v error \
-show_format \
-show_streams \
-show_entries stream_tags:format_tags \
-of json input.mp4
That command gives you the mechanical facts. It does not verify authorship or edit history.
For provenance, extract and validate C2PA separately, then merge the result into the same metadata record used by the rest of the pipeline. In practice, that means one asset record can answer both media questions and trust questions: what codec is this, what was the original capture timestamp, and does the manifest still match the bytes after a transcode?
This matters even more once files move through automated systems. RenderIO, webhooks, and no-code tools need a stable schema, not ad hoc tag scraping. A good reference point is RenderIO's file metadata concepts guide. The key design choice is to treat metadata extraction as pipeline observability. Capture what is in the file, what changed during processing, what survived export, and what can still be verified. That is the difference between reading tags and running a video system that holds up in production.