The Song Factory Now Checks Its Own Work — and the First Audit Flagged 6 Songs Out of 20
Every night the M.U.S.I.C. World HQ Song Factory writes twenty songs, fires them, and lands them — five lead voices, four songs each, two variants apiece. Forty clips before most people are awake. This week it started doing something it has never done before: listening back.
What changed
Until now the nightly quality gate was structural. Did every record land? Was Variety set to Off? Did anything trip moderation? Good checks, all of them — and all of them answerable without hearing a single note.
As of this morning's run the factory connects straight to the generation platform and pulls the finished audio back for measurement. Tempo and key, one clip per song, twenty out of twenty. For the first time the factory can compare what it asked for against what actually came out of the speaker.
What the first audit found
Fourteen songs came back on plan. Six did not, and they failed in patterns worth knowing:
Four landed at a clean multiple of the requested tempo. Two came back at roughly 1.5x — 78 BPM requested, 117.6 delivered; 80 requested, 119 delivered. Two came back at almost exactly double: 60 became 120, and 58 became 117.5.
Three slow trap records read as half-time — 138 measured as 69, 144 as 72, 134 as 67. Those are not errors. That is the genre working correctly, and a gate that flagged them as failures would be the thing that is broken.
Three came back in the relative major when the brief asked for minor. One asked for F minor and delivered A♭ major. One asked for G minor and delivered G♭ major. A third held its tempo within two BPM but arrived in B major instead of the minor key it was written for.
One was a plain miss — 90 BPM requested, 78.2 delivered, no genre convention to explain it.
Moderation flags: zero. All forty clips landed with correct titles inside a twelve-minute window. The batch was healthy. It just was not identical to the plan, and until this morning nobody could have told you that.
Why this matters more than the numbers
Here is the lesson, and it applies to every one of you working in Suno, Udio, Soundful or anything like them: the BPM and the key you type into a prompt are a request, not a setting. They are read as style guidance alongside everything else in the box. The model weighs them against genre, against groove, against the vocal it is building. Most of the time it honors them. Sometimes it rounds, doubles, halves, or quietly relocates your song to a neighboring key.
Tempo drift you will usually catch, because you can feel it. Key substitution is the quiet one. Minor to relative major is the smallest possible move on paper and the biggest one in the room — it is the difference between a record that aches and a record that shrugs. A song written to sit heavy will arrive sounding pleasant, and if you never check, you will assume you wrote it that way.
Three checks before you call a song finished
Pull the finished file into whatever you mix in and read its actual tempo and key before you build anything on top of it. Compare that against what you asked for, and write both numbers down. And when they disagree, decide on purpose — keep the version the model gave you because it is better, or re-fire with the brief tightened. Either answer is fine. Not knowing is the only bad outcome.
The factory vault has passed twelve hundred titles. Measuring the output is how the next twelve hundred get better instead of just bigger.
Published by M.U.S.I.C. World HQ | musicworldhq.com