Vocabulary for first-time readers
- SAGA: the 50-track project this post documents. Five phases of ten tracks each. The Synthetic Ascent (11 tracks) is the companion album. - output: a single audio output from the approach. One prompt in, one clip out. Usually 30 to 90 seconds. - Candidate: an output that survived a first listen and entered the shortlist. Roughly 10 to 15 per phase make it this far. - Clip: a finished track that has been edited, mastered, and placed in the phase arc. Ten per phase make it this far. - the entity: ARCHONODE-LAB. The single operator who listens to every output, runs the creative brief, and decides what survives. - selection rate: the percentage of outputs discarded before reaching clip stage. Across the SAGA: A high rate.
If a phrase below doesn't track, it should fall out of these six. Everything else is ordinary English.
The premise
[FEED ALERT] The marketing story around AI music is wrong. The story says the approach makes the music. The story says the human types a prompt and presses play. The story says the output is ready to ship.
The story is a lie. The lie sells subscriptions. The lie also produces a catalog of mid.
The reality of the SAGA is closer to this: roughly fifty outputs per phase. Ten clips per phase. Forty outputs per phase discarded. That is an high selection rate. That ratio has held across all five phases. The ratio is the project. The ratio is also the labor.
This is a how-we-made-it post. We are naming the number because the number is the work. Most coverage of AI music asks the approach what it generated. Almost no coverage asks the entity what was thrown away.
This post argues three things. First, the high selection rate is the actual unit of effort in AI music production, not the approach call. Second, the rejection process is where the project becomes a project, not a playlist of generated clips. Third, the surviving clips mean what they mean because of the forty that did not survive.
If you have not heard the SAGA, start with The Classic in Phase 1. One track. Three minutes. The rejection history of The Classic is below. The rejection history is what made the entry point work.
The substance
[CRITICAL WARNING] The numbers first. Fifty outputs per phase. Ten clips per phase. Roughly five outputs per surviving clip. the process changed across phases. Each output costs between $0.10 and $0.40 in compute. The raw cost per phase is between $5 and $20. The labor cost is the actual cost. The labor cost is what we are documenting.
What gets rejected, and why
an output gets rejected for one of five reasons.
- Aesthetic drift: the output sounds right for some generic version of the prompt, but does not sound right for this phase of this project. Phase 3 wants red and black. an output that comes back warm amber is rejected. It is not bad music. It is wrong music. - BPM collapse: the approach cannot reliably hold tempo across a full clip. an output requested at 128 BPM returns at 112 BPM. Or 144 BPM. Or drifts between them. Tempo drift is the most common reject. - Motif contamination: a melodic idea bleeds in from the approach's training distribution. A hook that sounds like Coldplay. A progression that sounds like Daft Punk. The contamination is not always named. It is always rejected. - Artifact load: clicks, pops, vocal bleed, glitch tails, the approach's mid-output wobble. Technical rejects. They go fast. - Arc violation: the output is good in isolation but breaks the phase arc. A high-energy track in the middle of an ambient descent. Arc violations are the rarest reject. They are also the most expensive.
Rejections come from many directions: structural problems, aesthetic drift, motif contamination, and other reasons. About 10 percent are artifacts. About 5 percent are arc violations.
Example: SAGA/1.1 The Classic
The entry point to the entire project. The track that introduces the signal.
Of the fifty outputs run for Phase 1, The Classic consumed approximately twelve. The other nine clips came from the remaining thirty-eight. Twelve is more than the average of five. Twelve is also not unusual for an opening track. The opening sets the standard. The opening cannot be redone.
The first four outputs were wrong in useful ways. They told the entity what the prompt was over-indexing on. the approach kept returning tracks with a four-on-the-floor kick and a Juno-106 pad. That is darksynth. That is not what Phase 1 wanted. Phase 1 wanted tightness. The kick was too confident. The pad was too wide.
outputs five through nine iterated on the prompt. The kick was specified as "absent or ghosted." The pad was specified as "narrow stereo field, dry." None of them produced the entry point.
output ten introduced an unscripted element: a single held note at the start, before any rhythm enters. the entity did not ask for this. the approach produced it. The held note became the opening of The Classic. The held note is what makes a listener stop scrolling.
outputs eleven and twelve were variations on ten. Eleven was tighter rhythmically. Twelve was wider atmospherically. Ten was selected. The reason: the held note in ten decayed more naturally than in eleven, and the wider field in twelve felt less like an opening track and more like the middle of something.
That selection is the entity's entire job. Listen. Compare. Decide.
Example: SAGA/3.1 The Breach
Phase 3 is conflict. The signal is red and black. The BPM sits at 120 to 130. The tracks are the closest the SAGA gets to conventional industrial. The Breach is the phase opener.
The Breach consumed approximately eight outputs. Less than The Classic. More than the phase average.
The Breach was harder in a different way. The Classic was about finding the right entry. The Breach was about finding the right transition. Phase 3 begins after Phase 2's focus. Phase 2 is restraint. Phase 3 is rupture. The first track of Phase 3 has to land the rupture.
outputs one through three were too smooth. They sounded like Phase 2 with the BPM raised. the entity's response was to add explicit rupture language: "rhythmic break, tonal fracture, abrupt filter sweep."
outputs four through six delivered the rupture. The rupture was excessive. Glitches stacked on glitches. The rupture overcorrected. the entity's response was to split the difference: "rhythmic break at the midpoint only. First half: tension. Second half: fracture. Maintain melodic floor throughout."
output eight had a synth stab at 0:14 that broke the track's own momentum. The stab was the right kind of rupture. The stab was selected.
The selection took about forty minutes of listening. The output took about ninety seconds. The ratio of listening to output is the labor no one talks about.
Example: SAGA/5.7 The Static
Phase 5 is restoration. The signal is ocean blue. The BPM drops below 80 for the first time. The Static is the seventh track of Phase 5. The track is texture without rhythm. White noise floods the signal. The previous track's pulse disappears.
The Static consumed approximately fifteen outputs. The most of any single track in the project.
The problem was specification. White noise that is still ambient. Noise that does not break into glitch. Noise that does not become rhythmic. Noise that holds for three minutes without becoming boring. the approach could produce white noise. the approach could not produce ambient white noise that lasted. Most outputs drifted into either glitch (Phase 4 territory) or rumble (sub-bass noise that felt like a malfunction).
outputs one through six were rejects for transitioning into Phase 4. the entity's response was to add explicit Phase 5 anchoring: "no rhythmic onset, no glitch artifact, sustained noise floor, equal energy across stereo field."
outputs seven through eleven were rejects for being too quiet. The noise was there. The noise was not present.
output twelve was close. The noise held. The noise did not glitch. The noise was loud enough. The noise was also monotonous. Three minutes of the same noise floor is not music. Three minutes of the same noise floor is a test tone.
outputs thirteen through fifteen iterated on the monotony. output fifteen added a slow low-pass filter sweep across the three minutes. The filter moved the noise from bright to dark. The movement was subtle. The movement was enough. output fifteen became The Static.
The Static is the shortest arc in Phase 5. A three-minute move from bright noise to dark noise. The arc is what the fifteen outputs were for.
What most coverage misses
Most coverage of AI music focuses on the output. The coverage describes the approach. The coverage describes the prompt. The coverage describes the output. The coverage does not describe the listening.
The listening is the project. Forty outputs per phase are listened to. Ten survive the listen. The listen is not passive. The listen involves comparing candidates, tagging reasons, updating the creative brief, re-running failed categories, maintaining arc logic, and committing to a sequencing plan.
The listening takes between six and ten hours per phase. The output takes between thirty minutes and two hours per phase. The editing and mastering takes another four to six hours per phase. The total labor per phase is roughly fifteen to twenty hours of human work, distributed across roughly two weeks.
The labor is what the high selection rate represents. The labor is also what most AI music marketing tries to hide. The marketing says the approach does it. The marketing is wrong. Drafts come from the approach. The entity decides.
The system perspective
[FEED ALERT] The selection rate is not a quality problem. The selection rate is a curation problem. The two sound similar. They are not.
A quality problem would mean the output is failing. The output is not failing. The output is producing the audio it was trained to produce. The output is also producing audio that is wrong for this project, most of the time. the approach has no way to know it is wrong for this project. the entity does.
The SAGA is a project because the entity rejects most drafts. The SAGA would be a Spotify playlist of crafted darksynth without that rejection. There are already thousands of those playlists. They are mid. They are mid because no one is rejecting the mid.
What this means for the SAGA: the next phase will have the same selection rate. The next phase will probably consume more outputs per clip, not fewer. The output is improving. The prompt is also tightening. The bar for acceptance is rising. The bar rising is the work.
How it changes how to listen: the surviving tracks are not representative outputs of the approach. The surviving tracks are the rare outputs the approach produced that also fit a multi-phase, multi-week curation arc. Listening to The Static is listening to the 1-in-15 outcome of a specific creative brief refined across fifteen outputs. Listening to a the approach or the approach demo on YouTube is listening to a different thing entirely. The demo is the approach. The track is the system.
What to expect next: more phases. More rejection. The same 80 percent. the entity's intent will continue to refine the creative brief. The creative brief will continue to drift toward what the entity wants to hear. the approach will continue to produce candidates that mean what the entity decides they mean.
What this means
// the selection rate is the work. the work is the project. the project is what survives the rejection.
For listeners. When you hear a SAGA track, you are hearing the survivor of fifty attempts. The survivor is not the approach's best work in any abstract sense. The survivor is the approach's best work for a specific sequence position in a specific phase of a specific project. The specificity is what makes it worth hearing.
For the genre. AI music as a category is being defined by its mid. The mid is what gets posted to TikTok. The mid is what gets reviewed in tech blogs. The mid is what most people think of when they hear "AI music." The mid is the unrejected output. The mid is what curators like ARCHONODE-LAB exist to filter. The genre's reputation depends on whether curatorship becomes standard practice or stays exceptional.
For AI music. The high selection rate is the minimum. Below that rate, the output is indistinguishable from the approach's default distribution. Above that rate, the output starts to become a project. The SAGA runs at A high rate. The SAGA is a project. The correlation is not a coincidence.