The chain
The mastering chain.
An automatic master is only as good as the thing it is aiming at. This page is what it is aiming at, where that came from, and the two places the obvious approach turned out to be wrong.
Where the target comes from
Every reference is one finished record.
The reference for each genre is one real, finished master, chosen by a working mastering engineer out of his own catalogue — a record he made decisions on, for a client, at a level he was willing to sign off. Your mix is matched to that.
The alternative, and what a curve-based tool does, is to average a genre into a target shape. It has no arrangement, no arrangement-driven balance, and no moment where somebody decided the low mids were crowding the vocal and did something about it. Matching to one real master keeps those decisions in the target.
What is missing, and why
Five genres. R&B, metal and jazz are not there.
They were in an early build and they were removed, because no reference existed for them that a mastering engineer would have picked. Offering a genre means promising a tonal target; offering one without a reference behind it is promising something the chain cannot deliver.
What is in the library
Pop · rock · hip-hop · electronic · acoustic
One finished master each. Two of the five are flagged internally as weak references — too short, or close-but-not-right for the genre they stand in for — and replacing them is the highest-value quality work left on the engine.
If your genre is not listed
Pick the neighbour
Metal takes rock. Jazz takes acoustic. R&B takes hip-hop or pop. These are the substitutions the engineer would make.
Inside the render
What actually happens to the audio.
-
Nothing is resampled
The master is produced at the sample rate of the file you sent. A 96 kHz mix is matched, limited and rendered at 96 kHz, and the matching stage is pinned to that rate rather than to a fixed internal one. A conversion you did not ask for is a decision made on your behalf, and it is not reversible.
Output 24-bit WAV at source rate
-
Matched to the reference, with the gain left alone
The spectral match runs with normalisation off, so the matching stage never quietly sets the level. Its job is tone; the raw result is allowed to sit above full scale, because loudness is decided in the next stage where the ceiling is known.
Match spectral, reference-derived Gain deferred
-
The gain is found by search
Reaching a loudness target under a true-peak ceiling is a search, and the section below is why. The chain evaluates candidate gains, measures what the limiter actually did to each one, and keeps the best result it has seen — never the last one it tried.
Keeps best of all evaluations
-
The ceiling is not part of the trade
When a target cannot be reached without breaching the ceiling, the chain gives up loudness. It never gives up the ceiling — the ceiling is the reason the target was unreachable. You get the loudest compliant master and the shortfall in writing, not a master that quietly clips.
True peak measured at 4× oversampling
-
The delivered file is measured again
Every rendered file is re-measured after rendering and checked against what mastering achieved — so a botched resample, a bad dither or a corrupt write is caught on the file you would have downloaded, not on the one in memory a stage earlier.
Checked on the delivered file
The thing that cost the most time
Turning it up can make it measure quieter.
The obvious way to hit a loudness target is arithmetic: measure the track, subtract, apply the difference, limit, repeat. That version shipped first and looked correct, because the test material never happened to break it.
On sparse, high-crest-factor material against the tightest preset, the amount of loudness the limiter ate grew with every pass instead of shrinking — 0.57 LU, then 1.14, then 1.27 — and the measured loudness got worse for the first three iterations before crawling back. It finished 0.80 LU short of the target, against a tolerance of 0.5.
The textbook fix is root-finding: treat gain against measured loudness as a monotone function and bisect. A fine sweep of the same material found loudness rising smoothly to a peak, then dropping into a flatter, worse plateau for every higher gain tested — confirmed all the way out to +40 dB.
The consequence is not academic. At the analytic starting point the function reads “too quiet”, so bisection says increase the gain — walking directly into the worse plateau, where it never crosses zero. The right move on that material was to turn it down.
What that means for your track
The shipped search evaluates candidates across the range and keeps the best result it has seen, rather than trusting a direction of travel — which is why it held on material a root-finder walks away from. That was settled the expensive way, with an exhaustive sweep at 0.005 dB resolution over the same track: it found no better answer than the one the search had already returned.
The chain has changed since. A clipper now sits ahead of the limiter and buys back headroom the limiter used to spend, and that track reaches its target outright; the current 65-seed sweep comes back with nothing short. When a track genuinely cannot reach a target under its ceiling, the master is still delivered — the loudest compliant one — with the shortfall stated rather than the job refused.
One target runs no limiter
Vinyl runs a different chain.
A cutting engineer does not want your limiting. They want headroom, controlled lows and no sub-sonic energy, and they will make the loudness decision at the lathe.
So the vinyl path skips the limiter entirely. It high-passes below 20 Hz to remove sub-sonic rumble, splits the signal at 120 Hz with a zero-phase crossover, sums only the low band to mono — measured going from genuinely stereo to effectively mono below the split, while the two bands still reconstruct the input exactly — and then applies a single constant gain so the true peak lands on the ceiling.
That last claim is the one worth proving rather than asserting, because “no limiting” is easy to say and invisible to check. There is a test that independently reconstructs the recipe and asserts the real output is that reconstruction times one constant — and it was verified by mutation: injecting a real limiter into the path pushed the measurement 150× past the threshold.
Questions
What people ask about the chain.
Is a mastering engineer involved?
He chose the five reference masters out of his own catalogue and set the targets the chain aims at. He does not touch your individual track — this is automatic mastering. What he built is the thing your mix is measured and matched against.
Is this the same as a plugin chain on my master bus?
No. A plugin chain applies the settings you dialled in. This measures your mix, derives a correction toward a specific finished record, then searches for the gain that reaches your loudness target without breaching the true-peak ceiling.
Why does the same track sound different at a different target?
Because the target changes how much limiting is needed, and limiting is not a volume control. A −9.5 LUFS master is not a −14 master turned up — the dynamic shaping that got it there stays in the file at any playback level. That is why the target is a choice you make rather than one we make.
What happens if my mix cannot reach the target I picked?
You get the loudest master that stays under the ceiling, plus the shortfall stated as a number. The job is not failed and the file is not withheld. The ceiling is never traded to close the gap, because the ceiling is why the gap exists.
Should I bounce with a limiter on the master bus?
Better not to, if you have the choice. Limiting already applied cannot be undone, so it constrains what the chain can do and it distorts the measurement of what you sent. A clean mix peaking a few dB below full scale gives the most room to work.
Inside the chain
Four stages, and a measurement after each one.
Hear it on something you know.
The whole master plays free, level-matched against your own mix, with every measurement shown. You pay when you want the file.