Case Study 1: Four Chords, Two Performances

A dissection. The session below is a teaching version of an experiment I've run many times with students and collaborators — the pianist is a composite and the exact numbers come from one representative pass, not a lab study. The pattern is the takeaway, and the last section shows you how to replicate it yourself in twenty minutes.

Here's the experiment. One chord progression — Am, F, C, G, one bar each, 96 BPM, the four-chord family that has launched more streams than any law of physics. The part is simple on purpose: each chord played twice per bar, beats one and three, with a single bass note underneath. Two versions go into the same sampled piano:

Clip A is typed. Every note starts exactly on its gridline. Every velocity is 100. Every chord's notes begin at the same sample. Every note lasts exactly two beats. No pedal. It is correct the way a spreadsheet is correct.

Clip B is played — by a working pianist, on a weighted MIDI controller, one take, no click anxiety, just "play it like it's the intro of something sad but hopeful." Then we open her take in the piano roll and read what her hands actually did.

That's the whole case study: an autopsy of the difference. Because here's what makes this experiment worth your time — every single thing she did is visible in the MIDI data, measurable in milliseconds and velocity points, and therefore programmable by you, tonight, with a mouse. Feel is not magic. Feel is data with intentions.

The Numbers on the Grid

For orientation: at 96 BPM a quarter note lasts 625 ms, an eighth 312.5 ms, a sixteenth just over 156 ms. The DAW's resolution is 960 ticks per quarter note, so one millisecond is about a tick and a half. Timing offsets below are in milliseconds relative to the gridline, negative = early, positive = late.

Clip A, bar 1 (Am), beat 1 — the spreadsheet:

Five notes, one instant, one loudness, one length. Multiply by eight chord hits and that's the entire clip. You already know what it sounds like — you heard it in Jaylen's bedroom in the chapter hook. Doorbell.

Clip B, bar 1 (Am), beat 1 — the player:

Read it slowly, because everything in this chapter is sitting in that little table.

The chord is not a chord — it's an event 25 milliseconds long. Her hand fell bottom-to-top: bass note fractionally early, then a roll upward, the melody note landing last and loudest, 25 ms after the first. Nothing about it sounds "rolled" at this width — you'd never say "arpeggio" — but the chord blooms instead of detonating. Twenty-five milliseconds is the difference between a door opening and a door slamming.

The velocities have a government. Top note 83 — the loudest, because it's the melody and her hand voiced it. Bass note 76 — second loudest, the foundation. Inner voices 64–68 — present, supportive, deliberately tucked. That's a 19-point spread inside one chord, organized by musical function. Random velocity variation between 64 and 83 would produce a different chord every bar and a drunk pianist overall; hers is the same hierarchy every time, with small variations inside it.

Nothing lasts its full value. Every note released between 1.79 and 1.92 beats — daylight before beat three, because her hand had to travel and her ear wanted the gap. The releases aren't simultaneous either; the chord decomposes in milliseconds the way it assembled.

The Shape Across Eight Bars

One chord proves the micro-anatomy. The macro-anatomy shows up when you chart all eight hits. Here's the summary — beat-1 and beat-3 hits per bar, averaged across each chord's notes:

Four patterns, each one a chapter lesson wearing data:

1. Beat three is always softer and later than beat one. Every single bar. That's meter, felt: the downbeat leads, the answer relaxes. It's also the "repetition never repeats" rule — the second hit of each chord is a response, not a photocopy.

2. The whole performance leans late — except where it doesn't. Average offsets sit +5 to +14 ms behind the grid: relaxed, balladic, sitting back in the chair. Then bar 4 — the G, the turnaround, the chord that points back home — arrives three milliseconds early, and noticeably louder. She leaned in. You can hear her wanting the next phrase. One early, loud chord out of eight is not an error; it's the plot.

3. There's a dynamic arc with a destination. Velocities dip slightly in bar 1's answer, rebuild through F, lift through C, and peak on the G — a four-bar crescendo nobody instructed her to play, because phrases go somewhere and she's a musician. Chart the average velocities and you can see the sentence shape: statement, softening, build, arrival.

4. The hierarchy never breaks. Across all eight hits, the top note rides 10–12 points above the chord average. The melody stays voiced through every dynamic — softest bar to loudest, the government holds.

And one lane Clip A doesn't have at all: CC64. Her pedal goes down a few dozen milliseconds after each beat-1 chord lands and lifts just as the next chord arrives — legato pedaling, connecting harmonies while clearing the old one's wash. Sixteen pedal events in eight bars. The piano breathes with her foot.

 velocity, top note per hit              timing lean per hit
 100┤                      ●                +15ms┤ ●     ●
    │                ●        ●                  │    ●     ●  ●
  85┤ ●        ●                              0ms┼─────────────────●──
    │    ●  ●     ●                              │ grid           ▲
  70┤                                        -5ms┤            the G,
    └─Am───F────C────G──                          └─Am───F───C───early
      the phrase has a destination                 the lean has ONE exception

Eight hits, two charts. An arc with an arrival, a lean with one deliberate violation. That's what "feel" looks like when you write it down.

Clip C: The Cautionary Tale

Because someone always asks: we made a third version. Clip A plus the randomize button — ±18 ms timing jitter, ±15 velocity jitter, the "humanize" preset.

On paper, Clip C's statistics resemble Clip B's: offsets in the teens, velocities spread across a similar range. Played back, it's unlistenable in a way Clip A isn't. The bass note lands late on one chord and early on the next; the top note is sometimes the quietest in the stack; the same chord lurches differently every bar with no relationship to the phrase. Listeners who shrugged at Clip A ("kinda stiff") physically winced at Clip C ("is something wrong with it?"). Clip A sounds like a machine. Clip C sounds like a machine failing — which your ear, prediction engine that it is, finds far more alarming.

Same deviation sizes. Zero structure. The entire chapter in one A/B/C: deviation is not the ingredient. Structured deviation is the ingredient. Intention is what survives the measurement.

Program It Yourself: The Replication Recipe

You don't need a pianist. Here's Clip B's anatomy as a twenty-minute checklist — this is exercise C1's answer key, so attempt the drill blind first:

1. Build Clip A honestly. Four chords, two hits per bar, five voices, flat 100s, gridlocked, full lengths. Bounce it; it's your control.
2. Roll every chord bottom-to-top across 15–25 ms — bass note 2–5 ms early, each upper voice 5–8 ms later than the one below, top note last.
3. Install the velocity government. Top note loudest (start ~83), bass ~7 below it, inner voices ~15–18 below. Keep the hierarchy on every hit.
4. Write the meter. Every beat-3 hit: 4–6 velocity points softer and ~5 ms later than its beat-1.
5. Write the phrase. Arc the velocities across the four bars — dip, rebuild, lift, peak on bar 4 — and let bar 4's downbeat land 2–4 ms early. One lean-in. Just one.
6. Release early. Pull every note end back to ~85–95% of full value, slightly staggered. Add CC64: down just after each beat-1 chord, up at each chord change.
7. Bounce and blind-test against Clip A at matched level, eyes closed, or better — on someone who doesn't make music. The vocabulary they use ("this one sounds like a person") is the grade.

Twenty minutes. No new notes. The ghost moves in.

What This Case Study Wants You to Keep

First: feel is legible. Open any MIDI performance you love — many DAWs let you drag a .mid in and inspect it — and the magic decomposes into offsets, hierarchies, arcs, and exceptions. Reading performances this way is transcription for producers, and it compounds: every part you autopsy adds moves to your hands.

Second: the sizes are small and the structure is everything. Nothing in Clip B strayed past 25 ms or 20 velocity points. Beginners humanize with a sledgehammer and conclude humanization sounds bad. The professional dose is homeopathic by comparison — and organized.

Third: one violation does the work of ten. The single early chord in bar 4 carries more "performance" than all the late ones combined, because it breaks an established pattern for a musical reason. Establish, then deviate. That's not just humanization — that's composition, and it's Chapter 16's arrangement logic in miniature.

Fourth, the honest limit: Clip B, programmed from the recipe, is still a ballad intro performance. Play the same progression as gospel, as house stabs, as neo-soul, and every number changes — direction of lean, width of rolls, where the government allows exceptions. The recipe isn't "the numbers." The recipe is listen to how a player serves this music, then write it down. Which is, you may notice, just Chapter 4 with a pencil.