Three hit records taught our mastering engine to do less

Some early rules flagged problems in records that had already been professionally mastered. We used eleven purchased lossless releases to find out which rules were making those mistakes.

The test was simple: run the scan on finished commercial releases and inspect every proposed correction. These files served as negative controls. They could expose a rule that mistook an unusual sound for damage, even though we did not have the original mixes or know every production decision.

bad guy: a bass-heavy balance

Billie Eilish’s bad guy measured 35.4% sub energy and roughly +34.6 dB sub versus mid. It also had a narrow tone near 13.77 kHz. An early rule could treat both as faults. The finished release gave us a concrete reason to question that rule: a bass-heavy balance and a sustained tone can be intentional.

We could use this file to reject a universal threshold. We could not use it to choose the right EQ for someone else’s mix, because we did not have its premaster. A reference master tells us what was released, not how to turn another recording into it.

Anti-Hero: flat peaks in a loud master

Taylor Swift’s Anti-Hero measured −8.7 LUFS, +1.5 dBTP and 6.2 LU LRA. It contained flat sample runs that could trigger de-clipping. But those runs alone did not justify reconstructing peaks in an already-loud master.

That finding led us to include source loudness in the repair decision. Already-loud material receives constant downward gain when needed for delivery. A flat-topped pattern on its own does not trigger automatic peak reconstruction in that path.

Too Sweet: low-mid energy needs context

Hozier’s Too Sweet measured −8.4 LUFS, −0.3 dBTP and 3.8 LU LRA, with a low-mid ratio of +12.4 dB. The older rule could call that buildup and suggest a cut. Yet a prominent bass part changes that same measurement without establishing that anything needs fixing.

We also measured 253 active ten-second blocks across the eleven releases. Differences between sections of the same finished records exceeded several proposed “bad mix” thresholds. A bass-led opening and a full chorus need different context; their average cannot tell us where an EQ cut belongs.

What changed after the tests

The initial rules proposed de-essing on seven of eleven masters, de-clipping on three and an acoustic AI-origin verdict on five. Later, the sibilance trigger fired on all four direct Suno exports and nine of eleven human releases. A narrow resonance appeared in one Suno file and five human masters. Those features overlapped too much to identify origin or establish a defect.

In the follow-up test, none of the eleven human controls received an origin verdict. Ambiguous tonal and repair controls stayed neutral. The chain measured 0.00 dB of six-band tonal change on all eleven files; where it lowered the delivery level, displayed LRA stayed unchanged.

What these eleven files taught us

The scan still reports sub energy, low-mid balance, resonance, width and clipping evidence. The change is in what happens next. When the evidence is ambiguous, the control stays manual so you can listen before deciding.

This was a small test set, mostly one song per production characteristic. It helped catch overconfident rules. It was not large enough to define how a genre should sound or replace comparisons between an actual mix and its approved master.