melchordgendaily · 01
← all reports

Report 01 · July 6, 2026

Building the instrument

The first day was about measurement. Before you can make a generator sound like one producer, you need a way to see how close it is (terms of art: methodology page). We built that instrument, used it to map his real taste (rhythm, keys, chords), caught a silent training bug with a chart, and by evening the data had predicted a fix his ear confirmed on its own. This is everything from that day, in full.

2,094
projects mapped
200k
chords decoded
55%
style match (diagnostic)
3 vs 2
b25: his-data melody model beat the old baseline
4
b27: first 4 rating, on chords from his own clips
01 / 06the instrument

Does it play like you?

The core question needs a number, so we score the model against the producer's own music as distribution overlap: for each musical trait, how much of the shared area the two histograms cover. 100% means the model's output is indistinguishable from his catalog on that trait; low means a gap. It is a diagnostic, not a target, and it points at exactly which lever to pull.

Style match by trait: register 73%, density 59%, fullness 64%, duration 25%. The per-trait numbers are the finding (three close, duration off by 2x); the 55% composite is an unweighted mean of four hand-chosen traits and carries false precision, kept only so later references resolve. Histogram-intersection overlap; higher is closer to his own distribution.

The same story in medians

The overlap bars come from comparing full distributions, but the medians make the gap concrete. Register and fullness sit close. Duration is the outlier: his leads are less than half as long as the model's notes.

traityoumodelread
register (MIDI)6762close
note duration (s)0.210.44model too long
onset density (/s)3.04.0model busier
polyphony (voices)32close
02 / 06the scorecard

Are you sounding like yourself yet?

The same four traits, read as a plain-English scorecard. This is the instrument's headline view: at a glance, which dimensions already sound like him and which one is off.

The verdict is unambiguous. Three of four axes are close. Duration is the one trait that misses badly, and everything the model does downstream (feeling too busy, too in your face) traces back to it. That is the lever for day two.

03 / 06the data

Keys, quantified

This is the day's deepest dig, and it caught a bug that would have quietly poisoned every generation. The producer writes almost everything on the white keys and pitches the whole project afterward with Ableton's pitch plugin. So the raw files look overwhelmingly like C.

81.7%
tracks written in C
18.3%
pitched off C
2.4× down
908 down / 384 up
7,066
musical tracks

Where the music actually lands

That 82%-C is the written shorthand, not the sound. Estimating the real tonal center from the notes (after the pitch offset) shows a genuine palette: C anchors, but the music spreads across E, A, D, G and the darker keys.

natural (as written)only reached by pitching
True sounding key across the 1,000 projects in the day-one estimate (the full corpus holds 1,485). C is the anchor, not the whole story.

The pitch shifts themselves

Of the 18% that move, the shifts lean down, toward darker, lower keys. This is his taste showing through the workflow: build in C for speed, then pull the whole thing down.

down (darker)up
Pitch-device transpose in semitones, zero excluded. The mass is on the down side.

Why it was a model risk, and the fix

Train on the raw written pitches and the model learns that the style lives in C, then drags every generation back toward it. Decoding the raw lead corpus showed exactly that: a spiky pitch-class profile piled on C's white notes. The fix is transposition augmentation, copying every clip into all twelve keys before training, which flattens the profile so the model learns melodic shape instead of a home key.

before (spiky, C-weighted)after (flat across 12)
Pitch-class share of the training data. Spread dropped from 3.9% to 2.08% (flat).

The retrain

On the key-flat data (148,768 sequences, 7 times the biased run), the model trained clean: a monotonic descent from 4.92 to 1.64. It absorbed his melodic shapes with no key anchor.

Key-agnostic retrain, cross-entropy loss over 37,100 steps.
04 / 06the color

Your harmonic color

We decoded 200,340 of his chords. He almost never plays a plain triad: 7.2% of 200,340 decoded chords. The color comes from extensions: 24.9% carry a 9th, 18.4% an 11th, 22.2% a 13th, voiced sparse and wide, mean 2.28 notes per chord. This is the quantitative target the chord model aims at.

Share of his chords carrying each interval. Triads are rare; extensions are the signature.
2.28
mean notes / chord
13.9
voicing spread (semitones)
7.2%
plain triads: the rarity

One caution the later days sharpen: this fingerprint is a selector, not a recipe. Forcing a chord to hit these stats by rule sounds odd, because the real chords carry voice-leading and context the averages miss. The number finds candidates; the ear picks.

05 / 06fingerprint

Minor, and around 118 BPM

Alongside his own productions, we fingerprinted 459 tracks he saved as references, the music he actually reaches for. This is the steering signal: not what to copy, but the center of gravity the generator should sit near. Two traits stand out.

Mode: minor gravity

His taste leans minor, 57% of the liked tracks. His own productions come out close to that too (about 52% minor by the same estimate), so the lean is real on both sides, the music he loves and the music he makes.

Tempo: a cluster around 118

Tempo centers near 118 BPM with a clear peak at 140, the future-beats and plugg pocket. Not a strict target, but a strong prior for where a generated loop should sit.

Tempo distribution across the 459 liked tracks. Mean 117.9, peak cluster at 140.
06 / 06convergence

The data predicted the ear

The style-match instrument flagged duration as the number-one gap in the morning. That evening, with no prompting from the chart, his ear said the same thing about b27: the melody was too far forward. Two independent lenses, the data and the ear, pointed at the identical fix.

That is a first concordance between instrument and ear: one trait (duration), one evening, one rater. [Calibrated 2026-07-11; this line originally read "the whole thesis working on day one", which one concordance does not support.] Training on him beat the generic model (b25 rated 3 against the old baseline's 2), his own chords earned the first 4 (b27), and the instrument called a real problem before the ear did. Day two starts on the duration lever.

Sources