Three hit records taught our mastering engine to do less

We measured legitimate commercial masters to catch false repairs. The numbers were real. Our interpretation of them was wrong.

An automatic mastering system can look impressive when every knob moves. It can also erase the reason a song works. I bought eleven lossless releases and used them as negative controls: if our engine called a deliberate choice “damage,” the rule had to lose—not the record.

bad guy broke the “extreme means wrong” shortcut

Billie Eilish’s bad guy measured 35.4% sub energy and roughly +34.6 dB sub versus mid. It also contained a legitimate narrow tone near 13.77 kHz. An early detector could interpret those numbers as excessive sub and suspicious resonance. In the record, they are part of a sparse, intentional production.

That single file could reject a universal threshold, but it could not tell us the right EQ move for another song. We do not possess its premaster. A finished reference is strong evidence against reckless repair and weak evidence for copying a curve.

Anti-Hero separated a hot master from a repair job

Taylor Swift’s Anti-Hero measured −8.7 LUFS, +1.5 dBTP and 6.2 LU LRA. Flat runs were present, so an early rule could propose de-clipping. But a hot commercial master is not the same state as a quiet raw export with accidental hard clipping.

The engine now uses source state as part of the decision. Already-hot material takes constant downward gain when delivery requires it. It does not automatically reconstruct peaks merely because the waveform contains a pattern associated with clipping.

Too Sweet showed why “mud” 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. An older whole-track rule could call that buildup. In a bass-led band production, the same measurement can describe weight and arrangement.

Across 253 active ten-second blocks from all eleven releases, section changes inside known-good masters were wider than many thresholds we had considered “bad mix” boundaries. A chorus, breakdown and bass-only opening do not live on one static target curve.

The audit got worse before the product got better

Before correction, automatic de-essing triggered on seven of eleven masters, de-clipping on three, and an acoustic AI-origin verdict on five. Later testing found the sibilance trigger on all four direct Suno exports and nine of eleven human releases. Narrow resonance appeared in one Suno file and five human masters. Those features did not separate origin or prove a defect.

After the changes, all eleven human controls return no origin verdict, every ambiguous tonal or repair control remains neutral, and the current chain makes 0.00 dB six-band tonal change to every one of them. Where delivery needs attenuation, LRA remains unchanged at displayed precision.

Restraint is a feature

This audit changed the product philosophy. PolishMyTrack still measures sub, low-mid balance, resonance, width and clipping evidence. It simply refuses to turn every observation into an automatic correction. A control at zero can mean “the evidence does not justify changing your music.”

That is also the subtle reason to use the service: not because it promises to transform every track, but because it shows its work and can leave a deliberate choice alone.