What could come next
Sixty-four ideas for this instrument, and none of them is promised. Eight left the list on 31 August 2026 by being built, played and written up, and nine more left it on 1 September — the watchdog, the phrase, the cellular automaton, per-step odds, the leaning generator, EXIT, banks as scenes, the modulator’s shape on the screen, and the screens’ brightness. They are on the Features page now. Twelve more were rewritten rather than removed, because something adjacent shipped and the entry had come to describe a world that no longer exists — each of those now says what exists and what is still missing, which is more useful than either keeping it whole or deleting it.
This page was regrouped on 31 August 2026. It used to be organised by when an idea arrived and which project suggested it, and it carried rank numbers. Both are gone. Where an idea came from does not help you decide whether you want it, and two rounds of numbering had reached the point where six numbers each named two different features. Ideas are grouped by what they would change about playing the instrument, and each carries a name.
The tables are also numbered 1 to 64, straight down the page, so that an idea can be pointed at in a sentence — “thirty-four” rather than “how sacred beat one is”. The number is a position on this page and nothing more: it is not a rank, not a queue, and not a promise, and it moves when an entry is built or removed. The name is the identity; the number is only a handle for today’s reading.
What this page is
Everything else in this guide describes an instrument you can play today. This page is the opposite: it is the idea list, kept in the open so that anyone forking the project can see what has been considered and how expensive it looked.
The list was assembled by reading the code rather than by wishing — every entry has a known mechanism behind it, and several were rejected as impossible before they reached this page. Ideas that turned out to be already built have been removed rather than quietly left in, and everything that IS built is listed on the Features page, so that nobody has to guess whether something is a gap or a feature. What was considered and rejected is written down too, in Considered and refused — grouped by the reason, because the reason is the part that transfers to anyone forking this.
Cost is S, M or L. S means a day or less and nothing new underneath. M means new state or a new part of the surface, still inside the shape the instrument already has. L means a new mechanism, a new page of the surface, or a question that cannot be answered without measuring the hardware first.
Impact is 1 to 5, and it means musical consequence — how much the thing you hear changes — not how clever the idea is.
Within each section the strongest ratio of impact to cost comes first. That ratio is where an argument starts, not where it ends: it knows nothing about what depends on what. The countdown ruler ranked twenty-second while nine of the entries above it could not exist without it, which is why it was the first thing built.
Ideas this instrument already has
Most lists of generative and DJ tools suggest the same things, and a good part of that list is already built here. That catalogue moved to its own page on 1 September 2026: it had grown past the point where it belonged inside an idea list, and a reader arriving with “does it do X?” should not have to read a page about what does not exist yet to find out that it does.
→ Features — every feature, in one list, each linked to the page that walks you through it.
Read the two together. This page is what is missing; that one is what is there; and the section at the end of this page is what will not be built at all, with the reason.
The instrument plays itself
The instrument can invent a bar and it can schedule a gesture. Between those two there is nothing that decides what the next eight bars should be. This is the section that would fill that gap — and the safety it needs, because a machine that arranges by itself must also be able to say when it has stopped.
| # | Idea | What it does | Cost | Impact |
|---|---|---|---|---|
| 1 | An arranger that mutes and unmutes | Exactly one part enters or leaves every eight bars, chosen by the machine — adding while the track is thin, stripping while it is crowded, and never bringing in a channel with nothing in it. The whole arrangement is which channels are silent, and the tension comes entirely from what is missing. | S | 5 |
| 2 | One cue, answered by every part | Every eight bars the machine draws a single instruction, and the bass, the hats, the percussion and the chords each answer that same instruction in their own idiom. Today the channels wander independently, which never sounds like a decision. | M | 5 |
| 3 | Auto-pilot | Arm a build and a drop to repeat on a sixteen or thirty-two bar cycle, and the instrument plays its own arrangement while you work on a sound. | S | 4 |
| 4 | The same bar tomorrow | Variation worked out from the seed and the bar number rather than from whatever has happened since. Half of this is already true and it is the half nobody expected: a voice’s line is a pure function of one saved integer, so the register does come back exactly. What does not exist is the bar number part — the counter is advanced by rewrites rather than derived from where you are, so nothing can be computed ahead, and the drums have no equivalent at all. | S | 4 |
| 5 | Generate three, keep the best | Compose a few candidate bars, score them, and play the best one. A cheap generator and a cheap judge beat an expensive generator, and the judge is the same one the constraint knob would use. | M | 4 |
| 6 | A style as odds rather than a fixed snapshot | The snapshot packs are frozen settings. This is the same pack expressed as distributions and hard rules, so a style becomes a region to wander inside rather than a single point. It ships, offline. tools/style-sampler.py samples and blends, including the rule that settings which cannot be averaged are taken whole from one parent — see Generating starting points of your own. What is left is the instrument: nothing in the driver can read a distribution, so a style is still a fixed point once it is on the rig. | L | 4 |
| 7 | A seed you can see and save | The premise here was wrong and it is worth saying so. A voice’s randomness is not gone forever once it moves — the register is one integer, it is saved, and it comes back exactly. What is missing is only the seeing: no page carries a seed, so a line you liked is recoverable by loading the snapshot and not by writing a number down. | S | 3 |
| 8 | Rules checked and repaired at every write | A named list of things that must be true of each kind of part — four-to-the-floor stays on the quarters, the bass stays off the one, open hats keep their distance — checked and corrected as the pattern is written instead of hoped for. | S | 3 |
| 9 | The night review | The instrument has written a structured line for every event since it was built, and nothing has ever read one back. A summary of a jam — what drifted, what was rerolled, which channels were muted longest, how far the tempo moved — is a tool over a file that already exists. | S | 3 |
Building, landing, breaking
Gestures you fire that resolve somewhere musical, four or eight or sixteen bars later, instead of happening the instant your finger moves. That rested on something the instrument did not have — a sense of which bar it was on. It has one now: the phrase clock and ARM, DROP, the CHANCE ramp and RISE are built and written up. What is below is what else could fire into that clock.
| # | Idea | What it does | Cost | Impact |
|---|---|---|---|---|
| 10 | The build ladder | One number from 0 to 100 that moves hits, density, odds, ratchet and cutoff together. Turn it by hand, or arm it and let it climb over sixteen bars. | M | 5 |
| 11 | Tempo ramp | Six BPM over eight bars into the drop, then hold or snap back. | S | 4 |
| 12 | One-shot fills | A second pattern per channel, fired by a pad, played once, returning to the loop by itself. FILL is not this, though it is close enough to be worth separating: FILL adds density to the last bar of a phrase and returns on its own, but it is a knob rather than a pad, and it decorates the same line rather than swapping in another one. | S | 4 |
| 13 | TURN | The bar before any landing becomes a fill, and the original pattern comes straight back. Every build ends on a turnaround. | M | 4 |
| 14 | BRAKE | Hold a button and the tempo bends down; release and it returns at the next bar. A tape-stop that resolves. | M | 4 |
| 15 | Timed transpose | Up a fifth for eight bars, and back, with the return already paid for. | M | 4 |
| 16 | Sampled risers and impacts | A sample scheduled backwards from the boundary so its tail lands on the drop. The player for it is already running and idle. Not the shipped riser, which is a one-shot modulator sweep and carries no audio. | M | 4 |
| 17 | Velocity swell | The pattern leans harder across a build and settles after the drop. | S | 3 |
| 18 | Silence before the drop | The bar before a landing goes quiet on the parts you nominate, and everything else plays straight through, so the drop hits harder. BREAK at length 1 already makes this sound; what it does not do is the arrangement half — it fires under your finger instead of being anchored to a landing, and it mutes mixer strips rather than nominating parts. | S | 3 |
| 19 | Retrograde | The bar plays backwards for N bars and then unmirrors. | M | 3 |
| 20 | Count-in | Once a drop is armed, the last bar clicks, so you can play into it instead of counting. | S | 2 |
| 21 | Loop roll | Hold a button and the channel loops the last eight, four, two or one steps around the playhead, then rejoins. The beat repeat on STEP ▶ is the neighbouring answer: a fixed one beat, from the pattern’s start, on every generated channel at once. It sidesteps the timing question rather than answering it — a selectable window around where the playhead actually is is still the hard part, and still not built. | L | 5 |
| 22 | Note repeat | Hold a pad and it retriggers at a chosen rate — the Maschine gesture this instrument does not have. Measured, and mostly refused: the driver’s heartbeat leaves a quarter-note repeat about seven per cent out and a sixteenth about twenty-eight, so only the slowest rate could be offered honestly, and that is not what a player means by note repeat. Left here as an open question rather than a plan. | L | 2 |
How a pattern is made
The generator owns the pattern, and everything here changes what it is allowed to write — either by giving it a new way to invent, or by giving you a way to overrule it that survives the next bar.
| # | Idea | What it does | Cost | Impact |
|---|---|---|---|---|
| 23 | How much the machine may move a channel | One number per channel, from nothing to everything. At zero the channel holds its pattern rigidly bar to bar, so the tops can move while the kick stays put. MOVE ships — see MOVE — and most of the caveat below has been closed since this entry was written: the ratchet, chance and timescale ticks, the gate collapse, FILL and the beat repeat all ask now. The drop still does not. The chord walker now holds when every voice is locked, which is as far as it can honestly go: it moves ROOT, and ROOT is one setting shared by all three voices, so it cannot be gated for one of them without giving each voice a key of its own. Lock one voice while its neighbours are open and the key still moves under it. What follows is the original entry. The generators themselves already ask before they write; what does not ask is the bar-rate gestures — the ratchet, chance and timescale ticks, the drop, and the chord walker moving the key — each of which walks every channel whether or not you had an opinion about that one. | S | 5 |
| 24 | A hole the generator may not fill | This makes a step the machine is not allowed to fill back in — the difference between what the generator owns and what you own, at the level of a single step. The old premise is gone: a tap no longer wipes on the next turn of a knob, because it lives in a register that survives ROTATE and drift. But it is not protected either — HITS, DIVIDE and LENGTH reset it by design, and RHYTHM’s own evolution can fill it back in. | S | 4 |
| 25 | Bar conditions | A step that plays only on the third bar of four, or the last of eight. The instrument already counts bars, so the pattern stops repeating without anything moving. | S | 4 |
| 26 | Stay out of the kick’s way | A channel that quietly drops any hit landing on the kick, so two parts cannot fight for the same instant. Beat one may still double it, because that is the one place they should agree. | M | 4 |
| 27 | Reverse it every fourth bar | Transforms that apply to a whole channel on a period: play backwards every fourth bar, rotate a step every third, thin out one hit in five. Every control this instrument has is a value; these would be the first that are rules. | M | 4 |
| 28 | How far the generators may stray | One knob for how narrow the danceable lane is. At the top, a bar is measured before it is allowed out and the weakest hits are removed until it holds together; at the bottom you get the raw field. LANE ships on the drums — see LANE. On the three voices the knob is there and the arm returns without acting, so what is left is a voice’s answer to “does this bar hold together”, which is not the same question. | M | 3 |
| 29 | How sacred beat one is | One number: at the top, the first step of a pulse channel always survives drift, rerolls and the random register; below it, the floor is allowed to drop out. It removes the one thing the generative side can do that a dancer notices immediately. | S | 3 |
| 30 | Set a step’s pitch with two taps | Hold SHIFT, tap the step, tap the pad whose note you want. One grid, two meanings, no menu. | S | 3 |
| 31 | Accent and micro-timing | A velocity accent mask and a small timing nudge per channel: the weight that swing alone does not give. | M | 3 |
| 32 | A part derived from two others | A hat that is defined as “wherever the kick and the clap disagree”. It cannot collide with either, because it is made of them. | M | 3 |
| 33 | Euclid inside euclid | Split a hat pattern with a second euclidean figure, so the closed and open hats divide a line neither could make alone. | M | 3 |
| 34 | An undo ring for the rest | Reroll has a one-deep undo and a voice’s register has a four-deep one; LENGTH and the pattern knobs have none. This would give them the same ring. | M | 3 |
| 35 | Echo as real repeats | Every hit spawning actual repeats rather than an effect, each repeat able to move in pitch — pitch inside the feedback loop. Musically one of the strongest ideas on this page and the worst fit for how this instrument works, because writing repeats into a pattern is a pattern rewrite, and that is banned for a reason. Recorded rather than planned. | L | 3 |
Voices, sound and the mix
| # | Idea | What it does | Cost | Impact |
|---|---|---|---|---|
| 36 | Mixer scenes | A whole mixer picture — levels, pans, mutes — recalled at the next bar. Two or three make verse, chorus and drop. | M | 4 |
| 37 | Spare chains | A channel becomes a real synth rather than borrowing one, with somewhere to put it. | L | 4 |
| 38 | A different drum engine | Replacing the sampler with one that has a filter per drum — the only route to per-channel drum timbre that measurement has left open. | L | 4 |
| 39 | Sound changes land on the bar | A kit or preset change waits for the downbeat instead of arriving mid-beat, so sweeping through sounds stops leaving a seam. | S | 3 |
| 40 | Trim that follows density | Doubling a channel’s hits doubles its presence in the mix, and today you fix that by hand. This estimates how loud a part will be from the pattern as it is written, and trims it before you hear the problem. | M | 3 |
| 41 | Duplicate a channel | Copy a channel onto another group, pattern and settings and modulation, then let it diverge. | M | 3 |
| 42 | A character insert | A second effect slot per channel, for drive and colour rather than space. | M | 3 |
| 43 | A different lead synth | Surge XT on the factory snapshot’s LEAD chain, with its own parameter tables. It already ships on other chains in both snapshot packs; this is about the factory default. | M | 3 |
| 44 | Tempo wander | A fraction of a beat per minute of slow drift, to take off the last of the machine-grid feel. Small, and worth knowing that it deliberately gives up an exactness the factory snapshot was tuned for. | S | 2 |
| 45 | Choke groups | Closed hat kills open hat. Measured, and the sampler will not do it: a one-shot plays its sample to the end whether or not a note-off arrives. Any version of this has to cut the sound somewhere else. | L | 2 |
What the panel shows
Two screens of one-bit glass, sixteen pads and a ring of encoders. This instrument has one law it cannot break — a silent channel must say why — and most of this section is that law applied to things the machine now decides on its own.
| # | Idea | What it does | Cost | Impact |
|---|---|---|---|---|
| 46 | A view that stops computing what it hides | A display mode that does not merely hide the busy parts of the screen but stops working them out at all. On a rig whose one hard failure is a saturated connection to the controller, a view that removes work rather than pixels is a safety feature rather than a preference. | S | 3 |
| 47 | Curves and waveforms on the screens | The screens draw text, bars and now a modulator’s wave silhouette. The dirty-rectangle work this entry was waiting on has landed, so a patch of screen is affordable. What is left is the drawing itself: there is no line or per-pixel primitive, only whole glyphs and filled blocks. “A wave that moves” stays banned whatever the primitives allow — animating a value on these screens is the one thing measured to throw this controller off the bus. | L | 3 |
| 48 | A groove report | Every bar, a few numbers for what the pattern actually turned out to be: how syncopated, how dense, how much of it lands on the beat. It is the natural readout for the constraint knob and the weighted generator, which otherwise change a pattern with nothing on the surface explaining why. | M | 2 |
| 49 | Rhythms as a number you can say | Sixteen steps is four characters. A pattern becomes something you can read off a screen, write on paper, dial back in tomorrow, or tell another player over a phone. | S | 2 |
| 50 | Layout presets | Rearranging which knob sits where, saved per player. | M | 2 |
| 51 | Three focuses collapsed into one | A simplification of how the instrument decides what a knob is pointed at. The only entry on the list that would cost relearning. | M | 2 |
| 52 | A step stutter on the transport’s left arrow | Hold it and the machine stops advancing and repeats the step it is on; let go and it comes back on the bar. Its neighbour already holds the beat repeat, so the two form a ladder under one finger — one beat to the right, one step to the left — learnable in a gesture with nothing explained. The 1/16 roll into a drop is the most-played live gesture in this music and the instrument does not have it. | S | 5 |
| 53 | ERASE + pad, or nothing | Answered, and it kept its job — built, not yet played. The question was whether a gesture that only ever removes still has a meaning now that a bare tap both adds and removes. It does, and the meaning is ownership: on a pattern you recorded the notes are the truth and removing them is the erase, which is what it always did; on a pattern the machine writes, the register is the truth, so it clears the step there instead and the step stays gone. The original entry follows. A bare pad tap now both adds and removes a step, so ERASE + pad may have no job left. Worse, the two write to different places: ERASE deletes the note straight out of the pattern while a tap writes the register, so an erased step quietly returns at the next rewrite. | S | 3 |
| 54 | LENGTH means two things side by side | Answered by the second option — built, not yet played. The lens does not refuse the verb: refusing would take a control away, and the lens exists precisely because five spread pages were removed. It carries the units in its title instead — ALL LENGTH BEATS/BITS — and per channel the voice column reads LEN BITS where the drum keeps a plain LENGTH. The original entry follows.Under the lens, LENGTH lays pattern beats on the drums next to shift-register bits on the voices — two quantities that cannot be compared, in one row whose whole purpose is comparison. | S | 3 |
| 55 | The screens repaint whole | Done 2026-09-05 — built, not yet played. The screen is drawn in three bands now — the tabs, the page indicator and each of the four encoder columns — and each one erases only its own box and is sent only when its own contents changed. Turning one encoder costs about a seventh of what a repaint cost before, and selecting a channel about a fifth; a repaint where genuinely everything moved is slightly cheaper than the old one rather than worse. The original entry follows. Every repaint clears the framebuffer first, which forces a full-screen transfer: 2120 bytes per changed display where a single changed column would need about 300. Corrected 2026-09-01: the daemon half is exercised now — it sends per-screen dirty regions, and a flush costs eight reports where it used to cost sixteen. The remaining waste is at the driver end: it still opens every repaint by clearing the whole screen, and a clear discards the finer regions underneath it. It is the one path that has thrown the controller off the USB bus, and it is what makes “never animate a value on the screens” a rule rather than a preference. | M | 4 |
Recall, recording, and other machines
| # | Idea | What it does | Cost | Impact |
|---|---|---|---|---|
| 56 | Lock snapshots and morph | Store the whole state of the instrument to a pad; recall it instantly, or morph into it over sixteen bars with every continuous value gliding. The bank grid is the recall half, and it is narrower than it looks: a bank holds the eight channels’ pattern state and nothing else — no sounds, no globals, no modulators. No morph exists at all, and the morph is the interesting half. | L | 5 |
| 57 | An automation lane | Per-step automation stored inside the pattern itself: saved with the snapshot, played by the audio engine, immune to any stutter in the driver. | L | 5 |
| 58 | Blend several states at once | Hold two or three pads and hear a weighted mix of the states they hold, moving as you move the weights — rather than a fade from one to another. Settings that cannot meaningfully be averaged, like a scale or an engine, are taken whole from the strongest pad. | M | 4 |
| 59 | Punch-in | A take that starts on the next bar and stops itself after N bars, so both hands stay on the pads. | M | 4 |
| 60 | Control voltage out | Pitch and gates to a modular rig. Zynthian already has this — the page, the controls and the wiring support all exist — so it is about fifteen euros of parts and a setting, not a build. A precise analogue clock is a different and much harder thing; MIDI clock out of the Maschine’s own DIN socket is the answer there. | M | 4 |
| 61 | Import Maschine kits | Turn kits from Maschine’s own library into drum kits this instrument can load. The format gives up little more than a list of samples — which turns out to be all this rig needs, because it takes everything else from the channel. The converter could be shared; the sounds could not. | M | 4 |
| 62 | The DIN MIDI sockets | The Maschine has five-pin MIDI in and out on the back. They work, and this instrument has never used them. Not a build — a switch nobody flipped. | S | 3 |
| 63 | Genre tags for the snapshot packs | Someone has already published a curated table of which expansion belongs to which style, under a licence that permits reuse. The tags exist offline — in both pack manifests and in every snapshot’s filename. What is missing is the instrument end: a tag is never written into the .zss, so nothing on the surface can read one, and you find a style by knowing the number. | S | 2 |
| 64 | Sync with other gear over the network | Tempo and phase shared with a laptop or other machines on the same network, with no cable. Harder than it looks: the sequencer’s tempo is not exactly the tempo you asked for, so on a shared clock this instrument would be the one drifting — which is the problem such a sync exists to solve. | L | 2 |
Considered and refused
The ideas above survived a filter. This is what did not, grouped by the reason rather than by where it came from — because the reason is the part that transfers. If you are forking this instrument and one of these looks tempting, the paragraph under it is the argument you would have to beat.
It needs two decks, and this has one
Automatic transitions, crossfades, blends, isolator swaps, double drops, bridges, rewinds, beat-matching, tempo alignment between tracks, harmonic mixing by key wheel. A whole research literature exists here and almost none of it applies: this instrument has one deck and no tracks. There is nothing to transition from, and nothing to transition to. It generates the music rather than replaying it.
One idea did survive the crossing, and it is on the list above: a part that closes through a filter rather than being switched off. That is what a good transition actually is, applied to a single channel instead of a whole track.
It needs a model this hardware cannot run
Real-time generative audio models, latent-space blending between sources, anything that infers a waveform. The Pi already runs eight audio chains and a sequencer with a hard real-time deadline; there is nothing left for inference. And the interaction those systems are built around is typing a description of the music you want, on an instrument that runs headless and has no keyboard.
One thing did survive: their interface. Blending several weighted targets continuously, rather than fading from one preset to another, is a good idea independent of what does the generating — and it is on the list above as a pad gesture.
It needs typing
Live-coding environments are the most powerful generative tools that exist, and none of them can be adopted here for the same dull reason: this is a control surface with sixteen pads, nine encoders and two small screens, and the rig has no monitor attached. What can be taken from them is the algebra rather than the language — transforms that apply to a whole channel on a period, which is on the list above.
It needs a sensor that is not there
Systems that read a crowd, or react to what is happening in a game, are answering a question this instrument cannot ask. There is no audience input and no game. The nearest real signal would be a MIDI controller plugged into the back, which is a mapping decision rather than a feature.
It was measured, and the hardware said no
- Cutting one drum off with another. Closed hat killing open hat is standard on drum machines. The sampler here plays a one-shot to the end whether or not a note-off arrives — measured. Any version of this has to cut the sound somewhere else, which is why it stays on the list above at low value rather than being quietly dropped.
- Note repeat at useful rates. The driver's heartbeat leaves a quarter-note repeat about seven per cent out and a sixteenth about twenty-eight. Only the slowest rate could be offered honestly, and that is not what a player means by note repeat.
- A precise analogue master clock. The expansion bus this Pi has jitters by about a millisecond from user space — fine for pitch, and about eight tenths of a per cent of a sixteenth note at 125 BPM, which is wrong for anything trying to stay in time with it. Pitch and gates out are still worth doing and are on the list; the clock belongs on a MIDI socket instead.
And the honest arithmetic
Three surveys were run: thirty-one other projects built for this same controller, and about thirty-five generative and DJ music tools.
The controller survey found not one generative technique in the entire set — no probability, no evolving state, no humanisation. The music-tool survey yielded five ideas that were new to this list.
That is not a complaint about other people's work, which is frequently better than this at the things it set out to do. It is a note for whoever reads this next: the search is closer to saturated than it looks, and the remaining value is in building what is already on the list rather than in finding more of it.
What the list says
Read down the ratios and one shape appears: the cheapest strong ideas are all about restraint rather than invention. A number for how much the machine may move a channel, a step it is not allowed to fill, a rule that beat one survives, a part that closes instead of vanishing — four small pieces of work, and between them they answer the one complaint a generative instrument earns, which is that it will not leave anything alone. Two of those four have since been built, which is the best evidence the shape was real: MOVE and EXIT both shipped on 1 September, and both were cheap exactly as predicted.
The expensive entries are expensive for the same reason, and it is not the code. Each needs a new page of the surface or a new answer about timing. On an instrument with sixteen pads and two small screens, where to put a thing is usually harder than how to build it.
And several entries are honest unknowns, kept here rather than quietly dropped. Whether a sampled drum can be cut off at all, whether the driver’s heartbeat is fast enough to roll a single step, and whether the screens will accept a small rectangle instead of a full-width strip are all questions for the hardware. The third has been answered since — yes, and it is in use, which halved what a display update costs. The other two are still open. This project has learned twice that a feature written from the code rather than from the rig is a feature that gets published wrong.