cut	file	what it is
v1	`out/cult-remix.mp3`	4:18, B minor, 128 BPM. Club banger. Kept for reference.
v2	`out/cult-remix-v2.mp3`	4:00, 120 BPM. Chill techno, and the words are sung.
v3	`out/cult-remix-v3.mp3`	Chorus-first: the dictated four-line chorus as the spine.
v4	`out/cult-remix-v4.mp3`	v2 re-scored — "run REAL fast", a harmonised "cult". 0:50 cut.
v5	`out/cult-remix-v5.mp3`	4:03. The whole record given a narrative, in ten acts.
v6	`out/cult-remix-v6.mp3`	4:03. v5's ten acts, with the polyrhythms rising as the voices sink and stair-step.
	`out/cult-remix-v6-50s.mp3`	0:48 → 1:38 — act III into the turn.
	`out/cult-remix-v6-extended.mp3`	0:48 → 3:28 — acts III through VII.

v6 — "the descent" #

@jeffrey heard v5 ("way better") and asked for three things:

as the track progresses can we bring up the polyrhythms a bit

and also bring down the vocals / so daaaaash daaaash gets lower in octaves

and can we like quantize it more like "bad" autotune it / eiffel 65 style

The acts, the narrative, the kick, the tube pump, the signal layer and the skids are all v5's and are unchanged. What v6 adds is that the three notes above are one move seen from either side: over four minutes the voices recede — quieter and literally lower, F#4 down to D3 — and the cross-rhythms rise into the room they vacate. It is a deliberate crossfade of attention. By act VII the lead is a stair-stepped machine near the bottom of its range and the thing carrying the record is five rhythms interlocking over one 4/4 kick.

bin/render5.mjs and bin/cut-v5.sh are untouched; v6 lives in bin/render6.mjs, bin/cut-v6.sh and bin/sing6.mjs.

1. The polyrhythm arc #

polyAmt(bar) is a single piecewise-linear 0→1 curve across the 120 bars. Five cross-rhythm layers switch themselves on as it crosses their own thresholds — nothing is scored per act; the arc is the score.

layer	period	enters at	made of
L1 · 3:4	4⁄3 beat (0.667 s)	act II, one tick	clicks; a closed hat from `a ≥ 0.36`; a ride from `a ≥ 0.66`; woodTaps wherever the machines are off
L2 · dotted eighth (3:16)	0.375 s	act III	taps — 5⅓ per bar, so the figure only re-meets the downbeat every three bars
L3 · 5:4	4⁄5 beat (0.4 s)	act V	clicks; every other bar until act VII, then every bar
L4 · 7:16	0.875 s	act VI	DTMF off the same C-U-L-T keypad, plus a bop — takes seven bars to come back round
L5 · the interlock skid	on the 3:4 grid	act VII	the TrackDrum friction voice, snare/rim band, cutoff 1900→1000, carrier 239→149

L2 and L4 are counted from t = 0, not from the downbeat. That is the whole point of them, and it is why the figure lands somewhere new every bar.

The arc, act by act:

act	`polyAmt`	hits/bar	what it is
I · carrier	—	0	nothing. A channel opening has no pulse, and the rise only reads as a rise if it starts from actual silence.
II · three voices	0.10 → 0.12	1.9	one 3:4 tick a bar, so the ear files it as colour
III · the message	0.28 → 0.34	7.0	3:4 every bar, and the dotted eighth begins to precess
IV · the secret	0.16	0.1 + woodTaps	the transmission has stopped: 3:4 in woodTaps only, no beeps
V · the reply	0.40 → 0.52	13.3	5:4 enters; 3:4 picks up a hat
VI · it spreads	0.62 → 0.68	15.0	the 7:16 phone figure joins
VII · the whole message	0.78 → 0.92	17.3	all five, interlocking, plus the skid on the 3:4 grid
VIII · recognition	0.80 → 0.55	12.9	the polyrhythms are the last thing to peel off, because by now they are what you are listening to
IX · the humans	0.18	0.1 + woodTaps	under the raw recording, a hand keeping 3:4 and nothing else
X · carrier off	0.30 → 0	2.6	the beeps outlast the music, as v5's did

Why it stays chill techno. The kick never leaves 4/4 and the hats never leave straight eighths. A polyrhythm is hypnotic when the pulse is unambiguous and the pattern disagrees with it; it becomes chaos when the pulse itself is negotiable. So the 4/4 grid is never touched and the disagreement lives entirely in voices small enough to be texture. No drops, no risers, no crashes — v5's rule, unchanged.

The levels were set by measurement, not taste. The first pass sat 25 dB under the bed (poly-only signal stem at −46 dBFS RMS against music at −16.8) and simply was not a cross-rhythm you could follow. The shipped levels put the layer where the record's existing phone voices already live — the narrative clicks run 0.44–0.70, the taps 0.38–0.68, the dials 0.46 — so the polyrhythm is as present as the phone and no more:

act	poly-only signal stem, RMS	full signal stem	music stem
II	−55.1 dB	−42.4	−21.0
III	−39.9	−36.4	−16.8
V	−37.4	−27.6	−16.6
VI	−35.5	−28.8	−16.7
VII	−33.0	−27.5	−16.6
VIII	−37.1	−34.5	−18.1

RMS under-reads a 4 ms click badly, so the honest companion figure is peak: in act VII the poly-only signal stem peaks at −1.6 dBFS, against the drums stem's −2.1. Per transient it is fully present; per second it is texture. That is the definition of a tick layer.

Debug: MUTE=poly3,poly5,polydot node pop/cult/bin/render6.mjs subtracts the whole arc; ONLY=poly3 renders one layer alone. (v6 also added the allow() gate to noiseHit, which was the one voice family MUTE= could not previously reach.)

2. The descent #

Two separate moves under one note.

The ladder. Every dash in the score names a rung — fs4, d4, b3 — and rung() walks that name down by however many steps the narrative has taken by this bar. The same written figure sinks an octave and a fourth between act III and act VIII without a single note being rewritten.

stage	from	`fs4` →	`d4` →	`b3` →
0	bar 0 (acts I–IV)	fs4	d4	b3
1	bar 48 (act V)	d4	b3	fs3
2	bar 64 (acts VI, VII/1)	b3	fs3	d3
3	bar 84 (acts VII/2–X)	fs3	d3	d3

The rungs below B3 are real lower takes of Camille and Alex, rendered by bin/sing6.mjs — not the same take pitched down. A transposed voice keeps the formants of the octave above it and reads as a machine, and the machine in this record is supposed to be the autotune, not the descent. Jeffrey already had the floor of the ladder from v5.

The ride. VOXG drops 1.42 → 1.22 and TUBEG 1.00 → 0.92 (a static −1.3 dB and −0.7 dB), and on top of that both vocal buses ride a piecewise-linear curve:

bar	0	24	48	64	76	88–96	104	120
gain	1.00	1.00	0.94	0.90	0.86	0.83	0.96	0.90
dB	0	0	−0.5	−0.9	−1.3	−1.6	−0.4	−0.9

It bottoms exactly where polyAmt peaks, and comes back up for act IX because the humans are the reveal of the whole record and must not arrive shy. The steepest segment moves 0.08 over 12 bars — 0.3 dB every 8 seconds — so there is no step anywhere in it, and the side return is scaled by the same curve so the space recedes with the voice rather than outliving it.

The tube bus deliberately takes less static trim than the vox bus: a held dash that is also descending into the pad's register has already lost presence before anything touches its fader. Measured on the act VII "a—waaaay" window, TUBEG 0.86 put it 2.6 dB under the bed; 0.92 keeps it where a lead belongs and still reads as quieter than v5.

3. Hard autotune — the Eiffel 65 stair-step #

bin/sing.py grew a real DSP mode, --autotune hard. It is not a preset name; it is the three settings that make "Blue" sound like "Blue", each pushed to its stop.

Snap. Every frame's target is forced onto the nearest member of a scale (B natural minor: B C# D E F# G A). Nothing lands between scale degrees, ever, so the pitch alphabet is seven letters wide.
Zero retune time. No portamento, no legato kernel, no overshoot, no preparation, no vibrato — the four Saitou expression effects that target_contour() exists to add are all skipped. The target jumps from one scale tone to the next inside a single 5 ms WORLD frame, and that discontinuity is the artifact you hear.
Rhythmic quantization. Frames are bucketed onto a note grid (125 ms = a sixteenth at 120 BPM; 62.5 ms for the short syllables) and each bucket is held at ONE pitch, the median of its own contour, so the steps land in time with the record rather than wherever the syllable happened to move.

What gets stepped is the performance's own pitch wander, not a sequenced melody: the source f0 is read along the same warp as the spectral envelope, high-passed against its own ~350 ms moving average (keeping the micro-wander, dropping the slow glide that is the melody's job), scaled by --autotune-drive, clamped by --autotune-range, and added to the scored note. That is why it reads as a retune of a real take.

Formants are never touched. cheaptrick's envelope goes into pw.synthesize() unmodified, so it is retuned, not chipmunked, and the words stay legible — which was the condition on the ask.

Two failure modes were found by measurement and fixed:

symptom	cause	fix
"run real fast" landing on A3 instead of D4	the whole slow pitch fall of a spoken word was being fed to the snapper	high-pass the contour against its own moving average (`--autotune-smooth-ms`)
"i wanna" jumping a fifth off the note	a 0.20 s syllable spans 1½ cells at 1/16, so one cell's median decided the whole word	finer grid (1/32) for short words and a leash (`--autotune-range`, 2–4 semitones) applied before the snap

Measured on run-real-fast at drive 3.2, pyin reports the notes at +8¢ / +13¢ / −2¢ — the melody is intact — while the frame-by-frame contour steps F♯4 → G4 → F♯4 → G4 → F♯4 → G4 → D4 → C♯4 → D4 → C♯4 → D4, every step landing on a 125 ms boundary. That is the sound.

bin/sing6.mjs renders the at- bank: the hook and chorus leads, and Camille's and Alex's dashes at every rung of the descent. Jeffrey and the dots stay human — he is the floor of the unison and the sub-octave under it, a stair-stepped bass warbles, and leaving one real person under the robot is what stops the stack reading as a synth. The crossover wiggle also halves where the autotune is on, because hand vibrato under a hard retune is two pitch modulators fighting and the retune wins anyway.

AUTOTUNE=0 node pop/cult/bin/render6.mjs renders the whole record with the human takes throughout, which is the A/B to reach for if the robot is ever too much.

Run it — v6 #

node pop/cult/bin/slice.mjs             # the sample bank (v1's, unchanged)
node pop/cult/bin/sing.mjs              # v5's speech-to-singing bank (cached)
node pop/cult/bin/sing6.mjs             # v6's descent + `at-` bank (cached; 26 renders)
node pop/cult/bin/render6.mjs --stems   # score → out/cult-remix-v6-full.wav + bus stems
bash pop/cult/bin/cut-v6.sh             # master once → full + 0:50 + 2:40 mp3s
node pop/cult/bin/qc.mjs pop/cult/out/cult-remix-v6.mp3

cut-v6.sh is cut-v5.sh's method exactly: measure the render once, apply one static dB of gain, limit. Never a second loudnorm pass.

Measured — v6 #

check	full	0:50 cut	2:40 cut
duration	243.2 s	50.0 s	160.0 s
integrated loudness	−14.0 LUFS	−12.5	−13.3
true peak	−1.3 dBFS	−1.7	−1.6
loudness range	8.2 LU	3.1	3.2
max raw sample step	0.520	0.357	0.514
max 2 kHz-band step	0.135	0.112	0.133
full-scale runs ≥ 3 samples	0	0	0
mono fold-down, whole track	−0.11 dB	−0.19	−0.10
mono fold-down, worst window	−0.86 dB	−0.68	−0.68

The raw-step figure is up on v5's 0.365 for the same reason v5's was up on v4's: there are more clicks in it on purpose — about a thousand more. Muting the three polyrhythm layers drops the count of steps > 0.35 from 27 to 11, which identifies them as the click voices themselves. The 2 kHz-band step is what actually distinguishes a discontinuity from a fast transient (a real signal in that band cannot move more than ~0.26 per sample), and it is 0.135 against v5's 0.117 — the same order, nowhere near the bound. No discontinuities were introduced.

Bus balance, from --stems (this is the check every earlier version failed at least once, and v6 is the version where it needed the most care, because the vocals are being pulled down on purpose):

bus	v5	v6	Δ
vox (sung words, dots, choir, grains)	−13.9	−15.9	−2.0
tube (the held dashes)	−16.5	−18.2	−1.7
drums (kick, hats, perc, skids, poly hats/ride)	−15.9	−16.0	—
music (bed, pad, stabs, bass, delay)	−17.5	−17.5	—
signal (beeps, bops, clicks, taps, poly ticks)	−25.4	−25.6	—

Whole-file integrated loudness hides what matters, so the check that actually answers "is the lead still proud of the bed" is measured on the phrase, over the 2.4 s a lead line occupies:

window	v5 vox	v6 vox	music	v6 lead over bed
"run real fast", act III (0:52)	−12.3	−13.8	−16.5	+2.6 dB
"run real fast", act VII (2:32)	−10.8	−14.3	−16.4	+2.1 dB

Two dB rather than v5's four-to-five. That is the ask, and it is still the loudest thing in the window.

v5 — "the signal" #

@jeffrey's note on v4 was five items, and the fifth reframes the other four: "lets add an overall narrative to the musical composition now".

The narrative #

The source is eight seconds of three friends chanting morse at a camera, captioned "dot dot dot dash dash dash dot dot dot (jk)" — SOS, disclaimed. So the story is already in the material:

A signal goes out, and it is answered, and the joke turns out to be true.

The arrangement spends four minutes taking the (jk) away. Ten acts, and nothing in the track exists that does not belong to one of them.

act	at	what happens	what the listener learns
I · CARRIER	0:00	DTMF tones dial C-U-L-T (keypad 2-8-5-8) into empty air over a drone. No drums, no words.	there is a channel
II · THREE VOICES	0:16	The pulse arrives and the three registers enter one at a time — Jeffrey B2, Alex F#3, Camille C#4. At bar 14 they land on one syllable for the first time and wiggle into unison. Then SOS, twice.	there are three people, and the signal is SOS
III · THE MESSAGE	0:48	The hook, legible, over the thick kick. The dashes ring in a tube.	the message has words
IV · THE SECRET	1:20	THE TURN. Kick out. The beeps stop — nobody is transmitting — and one lone click is a receiver being picked up. The word arrives alone, harmonised: cult.	the word is "cult"
V · THE REPLY	1:36	The hook returns, and in the hole it leaves SOS comes back beeped, in DTMF, from the far side of the field.	something answered
VI · IT SPREADS	2:08	No hook. Held dashes at the widest wiggle, a switchboard of beeps, and other people's "cult" — six utterances harvested from six other whistlegraph posts across 2022, arriving from the edges.	it has spread, and it no longer agrees about pitch
VII · THE WHOLE MESSAGE	2:32	The complete four-line chorus the record has withheld — run real fast / i wanna hide away / i wanna dash / dot dot dash — and the harmony stops wandering and comes home to Bm·D·G·Em.	the message, intact
VIII · RECOGNITION	3:12	Elements peel off, one per bar.
IX · THE HUMANS	3:28	Drums out. The original recording, unprocessed — plus somebody else, months later, singing it.	it was three friends and one take all along
X · CARRIER OFF	3:44	The drone leaves, the beeps outlast it, and the last sound is a hang-up click after the music has stopped.	the channel closed; the signal is still out there

Everything below serves that table.

1. Pitch wiggle at the crossover #

"when the dashes cross over we could definitely pitch wiggle with them"

dashStack() puts three real people on one syllable 28 ms apart, and act VI stacks two held dashes on top of each other. Those are the crossovers, and only there does the wiggle turn on — it is deliberately not a global vibrato. Each performer gets their own vibrato rate and phase (4.3 / 5.9 / 3.4 Hz, mutually irrational-ish so the beating never settles into a pattern the ear can predict) plus a slow detune drift in the opposite direction from their neighbour.

Two details make it work rather than just wobble:

the depth ramps in over the first third of each note, so the consonant lands in tune and only the held vowel shimmers. Intelligibility survives the effect;
the depth grows with the story — 4 cents when the three of them first meet in act II, 10 through the message, 16 once it has been answered, 22 where it spreads, back to 9 for the last statements. The narrative is a unison coming apart, so the wiggle is that happening.

It is implemented in the sample player: pos += step · 2^(cents/1200) per sample, which is a real varispeed read, not a post-hoc modulation.

2. The tubular pump #

"and pumping them with more tubular vibes"

The held dashes get their own bus, and it is the only thing on it:

a tube. Four feedback combs tuned to B2 · B3 · F#4 · D5 — the pitches the dashes and the Bm harmony actually live on — damped at 2.4 kHz in the loop. A tube is a delay line with feedback; that is physics, not metaphor. Each comb is scaled by (1-fb) so it resonates at unity instead of screaming, and the bank is mixed in parallel at 0.26, so the dry dash is still the thing you hear and the ring is what surrounds it.
tube saturation. tanh(1.35·(x + 0.11x²)), then a DC blocker. The x² term is what adds the even harmonics that read as tube rather than fuzz; the DC blocker stops an asymmetric shaper parking an offset on the bus and eating headroom. Drive 1.35 is colour, not distortion.
the pump. Depth 0.72, keyed off kick AND snare — against the bed's 0.50 — with a 9 ms ramp and a 0.26 s smoothstep recovery. A quiet rim on 3 exists mostly so the pump has a second trigger, which is what makes it read as rhythm rather than as the kick's shadow.

Hollow and ringing. No presence or air boost anywhere in the render.

3. A thick electro kick #

"want a thicker kick / POW / like a nice electro synth keek"

v4's kick was deliberately soft ("reads as a pulse rather than a punch"). v5 goes the other way, in five parts:

part	what
sweep	200 Hz → 47 Hz at 62/s, so the drop happens inside 20 ms and reads as one event rather than a bloop
envelope	two-stage — a 34/s slam over a 7/s tail — which is what makes a kick sound big and short at once
saturation	`tanh` at drive 2.4 on the kick's own waveform, folding in the harmonics that let a sine kick cut on a phone. This is the only tanh in the render and it is nowhere near the master
transient	two decaying blips at 1.6 k and 3.9 kHz, ~4 ms — there is now an attack to hear
sub	a separate 44 Hz layer with a slow decay, for the weight underneath

The POW is per hit. There is still no drop, no riser, no crash, no snare-roll build anywhere in the record.

4. Beeps, bops, clicks and taps #

"and more phone beeps and bops / and clicks and taps"

A fifth bus, ducked at half depth so it sits inside the groove rather than on top of it. A beep reads as a phone when it is two tones and not one, so these are real DTMF pairs off the keypad — and the record dials C-U-L-T, which is 2-8-5-8. Plus UI bops that drop a fifth in 60 ms, 0.9 ms clicks and short woody taps.

Placed narratively, never as a wall:

act I dials;
acts II–III tick under the voices;
act IV goes completely silent — the transmission stopping is the loudest thing in the record;
act V beeps SOS back;
act VI becomes a 3-against-4 switchboard that never lines up with the bar;
act X has the last sound on the record.

"menu band's trackdrum sliding sound could be GREAT to add — some skidding and sliding in the perc"

A direct port of the "Continuous membrane friction" block in slab/menuband/Sources/MenuBand/MenuBandPercussion.swift (~1296–1338), not an approximation. The parts that matter, kept exactly:

two low-pass states over the same white sample (at cut, and at max(35, cut·0.18)), with the friction signal being their difference — a tight moving band, not broadband noise. That band is the skid;
a sine carrier at res, frequency-modulated by the band itself (1 + tanh(band·8)·0.055). That 5.5 % self-modulation is the whole reason it sounds like a finger dragging rather than like filtered noise;
nonlinear grip: gnarl = tanh(band·(5 + rough·5)), then gnarl·0.44 + carrier·(0.08 + |gnarl|·(0.42 + rough·0.30));
a level target through separate attack/release one-poles (0.0025 s physical, 0.006 s synthetic), so it swells and releases like a real drag instead of gating on. The gesture lives in that envelope.

v5 adds a sweep on cut and res across each gesture, because a hand that is sliding is also moving and a static filter reads as texture rather than motion. Both stay warm — cutoff 700–2600 Hz, carrier 90–420 Hz — and the side send is held at ≤0.34 so friction never smears the mono fold-down.

Three gestures, 49 of them in the record: drag swells into a downbeat, skid comes off the back of a hit, slide is a slow swell-and-fall under a transition. Act IV is the showcase: with the beeps silent, a hand on a drumhead is the only thing still moving — the analog counterweight to a dead signal layer, and the sound of three people still in the room.

Other people's cults #

"get other samples from the other tiktoks / we have many 'cult' tiktok takes"

True, and more so than expected. system/public/whistlegraph.org/posts.json tags twelve posts with the cult work, not one. The other eleven were pulled from the assets mirror, transcribed with whisper.cpp, burst-segmented and chopped into alt/samples/ — 81 excerpts, including ten fresh utterances of the word "cult" (and, checked honestly with librosa.pyin: no whistles at all in any of the eleven — these posts are chanted, spoken and sung, not whistled). alt/harvest.json is the receipt.

Six of the ten are used in act VI, one per bar, thrown to the edges — the same word said by different people across 2022, which is literally what "it spreads" means. The sung one (alt-71448-cult, 0.80 s, 298 Hz) and a sung "the three of us are in a" go to act IX, and act X's last word is somebody else's.

The score receipt #

out/cult-remix-v5.events.json is no longer just metadata. Every scored hit is pushed into a flat events[] array from inside the voice that makes it, so a video renderer can draw the exact same performance the ear hears — 3158 events, each with t, voice, bus, dur, gain, pan and, where it applies, midi/midis/hz, sample, wiggleCents, DTMF digit, skid shape/cut/res, and who (camille / alex / jeffrey) on every performer-specific vocal hit — because three people on one pitch 28 ms apart is this track's signature and it should be visible, not only audible. The receipt also carries the ten acts with absolute start/end seconds and the narrative line for each.

What the source gave us #

The 8-second clip says exactly one sentence:

Dash, dash, dash, dot, dot, dot. The three of us are in a cult.

Six chant hits, and an RMS/autocorrelation probe puts them in three distinct registers — one dash and one dot from each performer:

voice	measured	note
Jeffrey	~127 Hz	B2
Alex	~193 Hz	F#3
Camille	~245 Hz	C#4
"cult"	~247 Hz	B3

B / F# / C# is a B minor power chord with the ninth on top, and the chant sits at 120 BPM — which is why v2 runs at 120 and not v1's 128. At the source's own tempo nothing has to be dragged to fit.

TikTok blocks this IP for that video, so slice.mjs pulls the mp4 from the AC asset mirror that whistlegraph.org's own posts.json already points at. Same file, no substitution.

v2 — the four notes #

@jeffrey heard v1 and asked for four things. Each one is a section below.

1. "we should be extending their words — use our whole speech-to-singing pipeline" #

The headline ask, and the centre of v2. Every lead line in the track is a sung note, not a chopped speech hit.

bin/sing.py is the full Saitou (WASPAA 2007) recipe on the WORLD vocoder (pyworld) — a thicker sibling of pop/bin/pitchsnap_world.py, which replaces an f0 contour but does not lengthen phonemes. Saitou's point is that speech is missing three things: pitched melody, sustained vowels, and a singer's-formant ring. sing.py does all three:

harvest / stonemask → f0
cheaptrick          → spectral envelope   (= vowel identity, = who it is)
d4c                 → aperiodicity        (= breath, consonants)
  ↓ DURATION CONTROL — warp the frame axis so the VOWEL absorbs the hold
    and the consonants stay short. A spoken "dash" becomes "daaaash",
    not a dragged "d-a-a-a-sh". Weights are voiced × loud × 70 ms past
    the vowel onset, so nothing stretches before the vowel has started.
  ↓ F0 SUBSTITUTION — write the score's contour, don't shift the source's
    (shifting is why v1's pitched hits still had spoken prosody on top)
  ↓ SAITOU'S FOUR SUB-EFFECTS — 42¢ attack overshoot decaying over 200 ms,
    a 22¢ preparation dip before rising intervals, 5.4 Hz / 32¢ vibrato
    faded in after 260 ms on notes longer than 0.45 s, and ±7¢ of
    lowpassed fine fluctuation
  ↓ SINGER'S FORMANT — +3.2 dB gaussian at 2.8 kHz, in the envelope
    (Sundberg 1974). Deliberately modest: @jeffrey does not want presence.
synthesize
  ↓ re-impose the voiced/unvoiced structure in the time domain and
    composite the WARPED ORIGINAL back over the unvoiced regions, so
    "s"/"t"/"sh" stay real instead of being coloured by tone
  ↓ de-ess (5–9 kHz, sidechained), 14 ms attack / 110 ms release

Everything pop/SPEECH-TO-SINGING-V2.md flags is applied: f0 is continuous through unvoiced gaps before synthesis (§2 — the word-start stutter), cheaptrick's f0_floor is set per source (§2 — held-note ring; 55 Hz for Jeffrey's B2 dash, 110 Hz for Camille), unvoiced regions come from the source (§6 — harsh sibilants), and verification is librosa.pyin, not autocorrelation (§1 — octave errors).

One extra trick that isn't in either doc: a long hold on a frozen cheaptrick envelope rings like a sample loop, so the read head gets a ±2.2-frame shimmer at 0.85 Hz inside stretched regions. That restores the formant drift a real held vowel has.

bin/sing.mjs drives it — 33 renders, cached by an args hash. Every one verified back with pyin:

word	source	sung to	stretch	pyin
dash (Camille)	0.31 s	F#4 / D4, 1.50 s	4.8×	+8¢
dash (Alex)	0.34 s	F#4 / D4, 1.50 s	4.5×	+8¢
dash (Jeffrey)	0.49 s	F#4 / D4, 1.50 s	3.0×	+8¢
dash (Jeffrey, bass)	0.49 s	E2/G2/A2/B2, 2.00 s	4.0×	−2¢
cult	0.54 s	11 pitches B2→G4, 4.00 s	7.5×	+8¢
dot	0.30 s	B3/D4/F#3/A3, 0.22 s	0.7×	+8¢
"the three of us are in a"	2.17 s	a 6-note melody, 4.00 s	1.9×	−12…+28¢
"i wanna"	0.40 s	D4→E4 / B3→C#4	1.3×	+8…+28¢
"run it fast"	0.59 s	G4→F#4→D4 held 1.20 s	3.4×	−12…+18¢

Median absolute error 8 cents, worst 28. The residual is mostly the vibrato and overshoot, which are supposed to be there.

The choir is one word. Camille's sung "cult" (B3), rendered at eleven pitches and picked per chord — a whole section's pad out of a single syllable. The hook's dashes are one syllable sung by three people: Camille, Alex and Jeffrey's dashes are each rendered to the same pitch and stacked 28 ms apart, so the unison is real performers, not a detune.

2. "the orchestration is weird" · 3. "can we like rethink the whole track melody" #

Full rewrite. 120 BPM, 120 bars, 4:00. No drops, no snare-roll builds, no risers, no crashes — v1 had all four. Chill techno is subtraction and patience: the arrangement moves by adding or removing one element at a time over a kick that mostly just stays.

bars	time	section	what changes
0–8	0:00	air	sung "cult" drone alone. no drums.
8–16	0:16	pulse	kick fades in, offbeat sine-bump bass, sparse hats
16–24	0:32	morse	the SOS figure, twice
24–40	0:48	hookA	the hook ×4, alternating full and reduced
40–48	1:20	hollow	kick out. "the three of us are in a" sung — rising, then answered falling
48–64	1:36	hookB	hook ×4, SOS dots answering in the hole the hook leaves
64–76	2:08	drift	no hook at all. harmonic travel, held sung dashes as pads, dub stabs, ride
76–96	2:32	hookC	fullest: hook ×5, 16th hats, clap, open hats
96–104	3:12	descent	elements peel off
104–112	3:28	ebb	drums out, the original spoken hook and chant return
112–120	3:44	out	the drone walks out the way it walked in

Measured RMS: −19 dB in air, −10.6 through the hooks, −15.5 in hollow, −22 in out. LRA 6.8 LU — the shape is in the arrangement, not the limiter. The three hook sections measure the same level on purpose; what changes between them is texture, not volume, because that is the difference between hypnotic and drop-bait.

The melody. The hook is four bars now instead of two, because a held word needs the room:

0.00  dash ─────────── F#4, 1.50 s   (three performers, one pitch)
1.50  i wanna           D4 → E4
2.00  dash ─────────── D4,  1.50 s
3.50  i wanna           B3 → C#4
4.00  run it fast       G4 → F#4 → D4 (D4 held 1.20 s)
6.00  dot               B3
6.50  dot               F#3
7.00  ─── rest ───                    (the chill needs the hole)

The SOS figure. The caption is morse, so it's a real four-bar riff: three short sung notes (B3 F#3 D4), three long ones (D4 F#4 E4, 1.2 s each), three short (D4 A3 B3). It earns its place because long-versus-short is the contrast a chill techno lead lives on anyway — and because the words being held are the dashes and the words being clipped are the dots.

4. "lets use some of our chord progression experiments" #

The harmony is lifted from the repo's own research, not invented: PROGRESSIONS_CHILL in recap/bin/trance.mjs, the eight-row pool that pop/dance/cutezenwaltzi.md describes as the chill mode walking "an 8-progression pool ... so it loops far less and travels more harmonically". Four of its rows, transposed to B minor, two bars a chord:

bars	row	degrees	in B minor
0–7	`[0,5,2,6]`	i VI III VII	Bm G D A
8–15	`[0,6,3,5]`	i VII iv VI	Bm A Em G
16–23	`[0,4,5,3]`	i v VI iv	Bm F#m G Em
24–31	`[0,2,5,3]`	i III VI iv	Bm D G Em

A 32-bar cycle — a full minute at 120 — which is long enough that the ear reads travel rather than loop. pop/dance/STUDY.md's rule that a dance track wants a simple looped minor progression is respected inside each row; the pool is what keeps four simple rows from becoming one boring one.

Mixing #

Follows pop/teknull/c/taksmukkeklokken-smooch.c, same as v1:

No clicks. Every voice exits through a 10 ms raised-cosine tail. The sidechain ramps to its floor over ~9 ms and recovers over 310 ms — and it is shallower than v1's (0.50 vs 0.66) because chill wants a breath, not a pump.
No clipping. No master tanh. Clean sum, linear trim to 0.92, loudness set only in mastering.
Sine bumps, not twang. Bass is fundamental + sub octave + a whisper of the 2nd through a one-pole, with a 75 ms glide. Pads and stabs are detuned harmonized sines. Nothing plucked, no presence or air boost anywhere — including inside sing.py, where the singer's formant is held to +3.2 dB.
Special Sign space. Dry equal-power pans keep it mono-safe; a cheap interaural model feeds a side bus band-limited 80 Hz–11.5 kHz and returned antisymmetrically (L=+, R=−) with a slewed send that breathes with the arrangement. Mono fold-down costs 0.12 dB across the track and never more than 0.82 dB in any 2-second window.
Dub delay — new for v2, and the one genuinely new mix element: a dotted-eighth (0.375 s) ping-pong with a 2.6 kHz damp in the loop and a 180 Hz highpass on the return, so the tails never muddy the sub. This is chill techno's actual reverb.
Four buses. Drums never duck, the harmonic bed takes the full sidechain, the sung voices take a quarter-depth duck (d^0.25) and ride +3 dB. That last number was measured, not guessed: --stems showed the lead sitting 6 dB under music + drums through the hooks, which is exactly the failure mode v1 hit. It now sits ~3 dB proud.

Mastering is measure → static gain → true-peak limiter, unchanged from v1 and for v1's reason: at this target loudnorm silently abandons linear=true and starts riding gain, which lifts the quiet intro ~20 dB and manufactures sample-to-sample steps the renderer was careful not to make (qc.mjs caught it at 8.79 s on v1).

Samples #

samples/ — cut by bin/slice.mjs from the whistlegraph itself and from the two ElevenLabs lines. Unchanged from v1.
sung/ — built by bin/sing.mjs from samples/. Derived; the .manifest.json (tracked) records what each word was sung to and what pyin measured back.
vox/ — the two ElevenLabs jeffrey-pvc renders, reused from v1, not regenerated. /api/say caches by content hash and v2 needed no new words, so this cut cost zero API calls.
pop/demos/samples/perc-*.mp3 — top-end percussion. The kick and all bass are synthesized here so the low end stays under control.
v1's whats-inside-your-heart stabs are dropped in v2. They were texture in a busy club mix; in a quiet one they read as the jokes @jeffrey warned they'd be.

Run it #

node pop/cult/bin/slice.mjs        # rebuild the sample bank (v1's, unchanged)
node pop/cult/bin/sing.mjs         # speech-to-singing bank (cached)
bash pop/cult/bin/render-v2.sh     # score + master → out/cult-remix-v2.*
node pop/cult/bin/qc.mjs pop/cult/out/cult-remix-v2.mp3
node pop/cult/bin/render2.mjs --stems   # bus stems, for balance checks

bin/sing.py needs pop/.venv:

python3 -m venv pop/.venv
pop/.venv/bin/pip install numpy soundfile pyworld librosa "setuptools<81"

(setuptools<81 is not optional — pyworld still imports pkg_resources, which setuptools 81+ removed.)

bash pop/cult/bin/render.sh still builds v1 and does not touch v2.

Measured — v2 #

check	value
duration	243.2 s (4:03 incl. tail)
integrated loudness	−14.1 LUFS
true peak	−2.0 dBFS
loudness range	6.8 LU
max sample-to-sample step	0.247 (0 steps > 0.35)
max 2 kHz-band step	0.122
full-scale runs ≥ 3 samples	0
mono fold-down	−0.12 dB overall, −0.82 dB worst window
pre-master max step	0.196 (0 > 0.35)
sung-note pitch error (pyin)	median 8¢, max 28¢

Run it — v5 #

node pop/cult/bin/slice.mjs             # the sample bank (v1's, unchanged)
node pop/cult/bin/sing.mjs              # speech-to-singing bank (cached; 93 renders)
node pop/cult/bin/render5.mjs --stems   # score → out/cult-remix-v5-full.wav + bus stems
bash pop/cult/bin/cut-v5.sh             # master once → full + 0:50 + 2:40 mp3s
node pop/cult/bin/qc.mjs pop/cult/out/cult-remix-v5.mp3

cut-v5.sh measures the render once, applies one static dB of gain, and limits. There is never a second loudnorm pass: at loud targets ffmpeg's loudnorm silently abandons linear=true and starts riding gain, which manufactures sample-step artifacts that the click scan then finds.

Different windows: START=80 DUR=40 SUFFIX=-secret bash pop/cult/bin/cut-v5.sh.

bin/render.mjs (v1), bin/render2.mjs (v2), bin/render3.mjs (v3) and bin/render4.mjs (v4) all still build their own outputs and are untouched.

Measured — v5 #

check	full	0:50 cut	2:40 cut
duration	243.2 s	50.0 s	160.0 s
integrated loudness	−14.0 LUFS	−12.3	−13.2
true peak	−1.7 dBFS	−1.7	−1.7
loudness range	8.2 LU	3.6	4.3
max raw sample step	0.365	0.249	0.249
max 2 kHz-band step	0.117	0.106	0.106
full-scale runs ≥ 3 samples	0	0	0
mono fold-down, whole track	−0.13 dB	−0.21	−0.12
mono fold-down, worst window	−0.79 dB	−0.64	−0.64

The raw-step figure is higher than v4's 0.186 for one reason: v5 has clicks in it on purpose. The 2 kHz-band step — which is what actually distinguishes a discontinuity from a fast transient, since a real signal in that band cannot move more than ~0.26 per sample — is 0.117 against v4's 0.125, i.e. slightly cleaner. No discontinuities were introduced.

Bus balance, measured from --stems (this is the check every earlier version failed at least once):

bus	integrated
vox (sung words, dots, choir, grains)	−13.5 LUFS
tube (the held dashes)	−13.8
drums (kick, hats, perc, skids)	−15.9
music (bed, pad, stabs, bass, delay)	−17.5
signal (beeps, bops, clicks, taps)	−25.4

The lead sits ~4 dB proud of the bed, which is the whole point of the measurement.

v10.1 — the words withheld, the dots everywhere, the phone made real #

Three asks from @jeffrey, one revision (bin/render10.mjs, gate: wordsIn()): "run real fast" and "i wanna hide away" are removed from every chorus statement and hook until act VII — before bar 76 the steady loop runs morse-only, with dot figures standing exactly where the worded lines will land. The shipped opening (bars 8–24) is now a dot collage from all nine harvested posts — each bar one video speaks a morse S into the dial's echo (raw alt/samples takes), answered by the aesthetivoxed sung/ dots in chord tones, both riding the signal layer's pans and delay so human dots and phone tones read as one transmission.

And the signal layer grew a real telephone: phone/ holds CC0 freesound gestures (receiver pickups, hang-ups, rotary dials, busy signals — ids + credits in phone/manifest.json, tracked). Rebuild after a fresh checkout with bash bin/phones.sh (decodes from the vault freesound cache; re-fetch via pop/bin/freesound-fetch.mjs if the cache is cold). The record now opens with a handset lifted under the first kick, act IV's "receiver picked up" click is an actual receiver, the act-VI switchboard has a rotary dial in it, and the last sound is a real hang-up with a distant US busy signal after it.

Second pass, same day, all from @jeffrey listening back: the opening collage was "too chaotic … wayyyy less of them … all extended a lot … more in the background" — so it became one dot at a time, each of the nine posts appearing exactly once, its longest take granular-stretched ~4.5× into a dark two-second sigh deep in the dial's echo. The arriving line lost its pickup ("no need for 'i wanna' before hide away") and got "sung much longerrrrr": sung/hideaway-hi stretched 1.9× across 3.8 s with an octave-down ghost and the tube "away"s moved late. And THE HUMANS act is cut ("remove the ending section at 3:00 where it's no longer sung … i don't like when vocals skip aesthetivox and feel too raw") — 120 bars → 112, the peel-off walks straight into CARRIER OFF, and no vocal on the record skips the aesthetivox anymore.

Shipped master: −14.0 LUFS integrated target, ~3:29.

pop / cult — "whistlegraph cult --- remix" #