the quality gate: real audio, real sync, real file

It takes almost no effort to generate a video these days. That is the problem. Autoretto runs on its own. It makes audio with Suno, artwork with Gemini, and motion with Sora. Then it renders a beat-synced video. If we let every render through, your channel would fill up with slop. So we built something between the render and the publish. We call it the quality gate. It is not a checklist written on a napkin. It is a series of checks that the final MP4 has to pass before we even think about uploading.

First, we actually decode the audio. We take the MP4 and run it through a real decoder. Not a quick header read. A full decode of the audio stream. We are listening for problems. We check for digital silence, for clipping, for corruption. If the decoder throws an error, the video is out. No second chance. The render gets flagged, and the pipeline stops. This catches files that look fine in a file browser but fall apart when something actually plays them.

Second, we verify the sync between audio and video. Autoretto's framework knows where the beats are because Suno can give us beat markers. But knowing is not the same as doing. The renderer might cut at the wrong time. So we extract the audio track from the final video and run our own beat detection on it. Then we pull the timestamps of every visual cut. We compare the two lists. If a cut drifts from a beat by more than four frames, we throw the video away. A few frames might not sound like much. On screen it looks like a flinch. Viewers notice that. We have to be strict.

Third, we refuse to claim a publish without a verified MP4 artifact. This is the simplest and most important part. The render process can say 'done' while the file is broken. A file could be zero bytes. It could be missing a moov atom. It could have the wrong duration. So we do not rely on the render report. We check the file itself. It has to exist. It has to have a non-zero size. We probe the container, we check the codecs, and we play it with a headless player. If playback fails, we treat it as a failed render. We will not upload something we have not seen play.

The gate runs before the scheduler even looks at the release. That is an important detail. The scheduler does not pick a video and then check it right before upload. It checks the artifact first. If the artifact is not verified, the slot is skipped. The channel stays quiet for that slot. We would rather miss a slot than publish garbage. This is also how we keep the performance data clean. Every published video has a real artifact behind it. When we look at watch time and retention, we know those numbers came from an actual file.

These checks are not about aesthetics. They are about honesty. A publish means the video is live on YouTube. It means your viewers can click play. If we claim a publish and the artifact is broken, we are lying. That is why the gate is strict. We hold the MP4 accountable for everything that comes after. The performance data only matters if the video actually exists.

That matters for the learning loop too. Autoretto improves the next release by studying what happened with the last one. If the last release never actually played, the signal is noise. A broken video is not a chance to learn. It is a failure to deliver. By refusing to publish without a verified file, we keep the feedback loop honest. The models only see success and real engagement. That is why we spend time on these checks instead of pushing more output.