It measures first. It never guesses.

What Burnish reads before it touches a sample.

The analysis runs first, every time. It's the difference between an engineer who listens before reaching for a fader and one who applies the same chain to everything.

• Integrated loudness (LUFS) and loudness range (LRA)
• True peak — 4× oversampled, catching the intersample peaks the waveform hides
• Crest factor, full-mix and per frequency band
• Five-band spectral profile — sub, bass, low-mid, presence, air
• Stereo width and correlation
• The AI artifact fingerprint — the HF cliff, mud accumulation, harshness signature
• PLR (peak-to-loudness ratio) as a dynamics health indicator

The difference between fixing and hurting.

This is the subtle part, and the most important. AI mud cleanup in the 200–400 Hz range overlaps directly with the chest resonance of a baritone vocal. A filter that clears muddiness can simultaneously thin a voice. A limiter chasing loudness can flatten the punch of a kick.

Burnish uses measurement to tell these cases apart. It checks whether a bass cut would remove sub-rumble or vocal body. It checks whether a loudness gain would push the crest factor past the point where transients crush. It re-measures spectral balance after every stage and backs off the moment the numbers move the wrong way.

That's what "adaptive" means — it responds to what it's working on, not what it assumed it would find.

You can't EQ in air that was never recorded.

When Burnish detects the hard cutoff that AI models produce around 14–16 kHz, it doesn't try to boost frequencies that aren't there — there's nothing to lift. Instead it synthesizes harmonics: new high-frequency content musically derived from the band below the gap, the way a harmonic exciter works, but aimed precisely at the missing region.

The result is air and shimmer where there was a ceiling — not the hiss and harshness a crude high-shelf would add.

A master that fails doesn't exist.

Every master clears nine release-grade checks before the file is written: true peak ≤ −1.0 dBTP, zero clipping, ≥ 60% of dynamic range preserved, ≤ 4 dB crest-factor loss, PLR in the 6–16 dB band, mono-safe correlation, no band shifted more than 3 dB, stable width, and loudness on target. Fail a critical one and the pipeline reverts and tries again, gentler.

Then it hands you the receipt. The report shows before/after for every metric, which stages fired and why, which were skipped, the gate results with measured values, forensic flags with timestamps, and a graded score across Loudness, True Peak, Dynamics, Spectral, Processing, Vocal Clarity, and Translation.

If a clarity score is low because the source has a 15 dB bass-to-presence gap, you'll see it — so you know the problem is the mix, not the master.