After three parts, the brief was to double four things. The reading. The openly licensed data. The claims we test. And the writing about where this goes. This page reports each doubling and what it changed.
The short version. The reading found a measure we had never taken. The data grew, and most numbers held. The claims got a register. Two of ours turned out to point the wrong way. The vision got dates.
Part one rested on 25 sources. A second sweep added 25 more. It went where the first was thin. How people hear ducking and deep bass. How they hear the pitch and grit of a kick. How machines find the downbeat in dance music.
It also covered three more things. How drum sounds are placed in learned spaces. What has been written about hard techno itself. And how loud masters are.
Three things came back. One is support. A 2023 study of groove agrees. Sound does most of the work, not timing.
One is a surprise. Producers describe kicks that double as their own bass. That is the sound our model failed on.
The third is a gap. None of our pages reported integrated loudness. We now do. The sweep also listed ten measures we could add. Two of them are cheap and will come next.
Part two searched four tags on the Internet Archive. The second round searched nineteen more. It also read every licensed techno item for certain words. Hard, industrial, rave and a few others. That gave 3,836 licensed tracks to choose from.
We fetched 444 more, again at most twelve per producer. The corpus now holds 508 kept tracks from 317 producers. Every one was measured the same way. Every one had a minute split into stems.
The real hit library grew with it. It now holds 4,138 kicks, 596 hats and 463 claps. All cut from the drum stems of real tracks. Across 399 tracks the kick lands at a median 55 Hz. Half sit between 48 and 66.
Most of part two's numbers held. Half the sub attacks land on the beats. Chance is a quarter.
Clear backbeat: 80 of 508 tracks. Pump on the mix: median 6 dB. Kick returns on the eight bar line: 23 in 100. The bigger net did not move them far.
Two warnings. The new tags reach past hard techno. Some new tracks are slower or softer than we meant. The tempo filter, 128 to 180, keeps out the worst. And the pool is still one archive. Free licence hard techno is a small world.
Integrated loudness is how loud a whole file is. It follows the broadcast rule. We measured it on the minute we split into stems. The stems were added back together first. The median is -10.7 LUFS. Half the tracks sit between -13.2 and -8.6.
The crest of the mix sits at 11.3 dB. That is quieter than a chart master. Those sit near minus eight. It fits a netlabel scene outside the loudness war. Loudness did not rise from 2018 to 2026 here. The crest did not fall.
Part one made its claims in running text. Now they live in a register. Each hypothesis has a measure and a test. Its prediction was written before the test. There are 42 of them. The data we have could test 35.
The count: 13 supported, 21 not supported, 1 reversed. A reversed claim points the other way. We had two. Both are about the same thing.
We predicted that tracks with a backbeat pump less. They pump more. With a backbeat, the median pump is 10.7 dB. Without, it is 5.5 dB. We predicted a separate rumble layer makes the kick brighter. The opposite holds: 1585 Hz against 2404 Hz.
Both predictions came from one picture of the music. Old tracks have a clap and a plain kick. New ones have no clap. Their kick is its own bass. The data say the picture is too simple. The backbeat tracks pump hardest of all.
Some claims held well. Pump depth falls a little as tempo rises. The sidechain release tracks the beat. The low end is back within two thirds of a beat. Tracks with a confident downbeat return on eight twice as often. And the kick body got a little brighter over the years.
The full table is in the register. It is built again from the data whenever the corpus grows. So are the results. A claim that fails at twice the tracks would say so.
A last sweep went after published advice. Producers, mastering engineers, the people who make the tools. It came back with six signal chains, stage by stage. Each stage names the measure it should move. It also brought 15 more claims.
We tested those the data allow. They are in the register too, from H33 on.
One result stands out. The tools promise a split duck. The sub dips deep and the low band stays. We looked for that split on the bass stem. It is not there. The sub and the low band dip by the same amount.
Whatever the tools do, the tracks do not show it. The climb back after the dip is slow, though. It takes about 72 ms to rise. The people who make the tools say they come back softly. This fits.
Some of that advice is folklore. It is repeated everywhere and measured nowhere. The register is where it meets the corpus.
Two new notes join the three. Timelines for an automatic mixing engineer puts dates on the agent. A month, a year, five years, ten. Each has a test you could run on that date. Each has the odds the first week gave it.
The second is A hypothetical mixing session with the finished system. It imagines one evening in 2029, written as a scene. It shows what it says, and what it must not. The notes are not plain language. They are letters to one producer. This page is the plain record of what they rest on.
The second sweep could not open three sources past the summary. One paper on ducking could not be opened at all. The new tags reach past the genre. The loudness number is from one minute per track. And the register tests what we thought to write down. The claims we did not think of are not in it.
research/literature-2.md. Advice sweep, chains and folklore: research/advice.md. Register and tests: research/hypotheses.md, research/test_hypotheses.py.scratch/discover2.py and corpus2/discover.py. Round two: corpus2/round2.py. Loudness: analysis/loudness.py, BS.1770 by way of pyloudnorm.