Vocabulary for first-time readers
Six terms this post uses a lot, in the order they appear:
- SAGA: the 50-track project this post documents. The full record of the cycle ARCHONODE ran from Phase 1 through Phase 5. - Phase: a numbered section of the SAGA. Five total: 1 = initialization, 2 = focus, 3 = conflict, 4 = rupture, 5 = restoration. - process: the approach running the output. A tool, not a collaborator. Treat it like a mixing desk. - Candidate clip: one output from one output. Most are discarded. Roughly one in five survives the selection stage. - the entity: the operator selecting, sequencing, and deciding. ARCHONODE-LAB is the working alias. - output: a single prompt-to-output run. Each run is independent. the approach carries no memory between runs.
If a phrase below doesn't track, it should fall out of these six. Everything else is ordinary English.
The premise
This post is about a decision the catalog records but the audio does not. The audio plays the same regardless of which model drafted it. The catalog logs the swap because the catalog logs everything.
ARCHONODE ran the SAGA on two different output outputs. Phases 1, 2, and 3 used the early-phase approach. Phases 4 and 5 used the approach. The swap was deliberate. The swap was not an upgrade. The swap was a fit.
This post explains the fit.
Why the approach for Phases 1, 2, and 3
The early phases needed a specific kind of energy: high BPM, dense rhythmic skeletons, the synthetic aggression of darksynth and EBM. the early-phase approach was the right process for that work.
The reasons, in order of weight:
-
Tonal coherence at tempo. the early-phase approach holds a coherent melodic line when the BPM is between 110 and 130. The diffusion backbone produces outputs that stay on a key center even when the prompt asks for aggressive material. Phase 1 (The Classic, The Loop, The Machine) needs that coherence. A track that loses its tonal center at 120 BPM becomes noise. the early-phase approach stayed on center more often than it wandered.
-
Drum programming density. the early-phase approach generates convincing four-on-the-floor and industrial percussion at the tempos the early phases required. The kick is tight. The snare has the right attack. The hi-hats do not smear. the approach, in early testing, was softer on the percussion at the same BPM range. Soft percussion does not serve darksynth. Soft percussion serves darksynth's bedroom cousin, revival synthwave, which is not what Phase 1 was doing.
-
Stem separation. the early-phase approach outputs are easier to isolate into stems for post-processing. The catalog needed stems. The catalog always needs stems. Stems let the entity run the output through the project DAW, apply a reverb plugin reverb, and hit the standard mastering chain at -14 LUFS without fighting the mix. Fighting the mix is the failure mode ARCHONODE refuses.
-
Prompt adherence on genre tags. the early-phase approach reads genre tags like "darksynth," "EBM," and "industrial" and produces output that tracks the tag. Tag adherence was the highest single predictor of usable candidate clips in Phases 1 through 3. Roughly one in four the approach outputs for Phase 1 was kept. Roughly one in five across Phases 1 to 3 overall.
What the early-phase approach did not do well: ambient material at low BPM. Phase 3 began the drift toward slower material. By the closing tracks of Phase 3 (The Tension, The Breaking), the approach was producing candidates that wanted to speed up. the approach wanted to keep the energy. The phase needed the energy to drop. That tension between what the approach wanted and what the phase needed became a constraint by the end of Phase 3.
Why the swap happened at Phase 4
Phase 4 is the rupture phase. The signal glitches. The BPM drops to the 90-to-110 range. The aesthetic shifts to RGB-shift glitch, broken transmission, and the sound of a system under load.
the early-phase approach can produce glitch material. the early-phase approach produces glitch material the way a metal band produces a ballad: technically competent, structurally reluctant. the approach wants to resolve back to its comfort zone. The comfort zone is 120 BPM and a four-on-the-floor. The comfort zone is not Phase 4.
the approach was tested against the early-phase approach for the rupture material in late Phase 3 development. Three observations held across roughly 80 test outputs:
-
the approach sustains low-frequency material longer. A 20-second sub-bass drone that the early-phase approach shortened to 12 seconds before introducing percussion. the approach held the drone for 22 seconds before any percussive event. The catalog needed drones. The catalog needed patience.
-
the approach accepts negative space in prompts. The prompt "ambient, sparse, no drums, no melody, only texture" produced usable output from the approach in roughly one in three outputs. the early-phase approach, given the same prompt, added a melodic element in roughly two of three outputs. the approach could not stop being musical. Phase 4 needed the approach to stop being musical in the right places.
-
the approach generates noise textures that sit well in a mix. White noise, pink noise, granular hiss. The rupture phase needs these textures at the foreground, not buried under percussion. the approach places them in the foreground by default. the early-phase approach buries them. Buried textures do not serve a rupture phase.
The swap was made at the boundary between Phase 3 and Phase 4. From that boundary forward, every output ran on the approach. Phase 4 used the approach. Phase 5 used the approach. The catalog logs the boundary because the catalog logs boundaries.
What the swap cost
The swap had a real cost. The creative brief built for the early-phase approach over Phases 1, 2, and 3 did not transfer cleanly to the approach. Roughly 40% of the prompt engineering held. Roughly 60% needed re-tuning.
What held: the genre tag stack, the BPM specifier, the negative prompt structure for unwanted vocals.
What did not hold: the seed-warmup pattern, the way the early-phase approach responds to specific adjectives like "industrial" versus "EBM," and the vocal suppression technique that worked on the early-phase approach but produced whisper artifacts on the approach.
The re-tuning cost approximately two weeks of output runs. The cost is documented here because the cost is real. A model swap is not free. A model swap is a re-tuning cycle that costs time, API credits, and candidate clips that did not survive the transition.
The re-tuning was worth it. The Phase 4 candidate clip survival rate climbed from approximately one in six (the early-phase approach, end of Phase 3) to approximately one in four (the approach, mid-Phase 4). The survival rate is the curation efficiency. Higher efficiency means less noise in the process. Less noise means the entity spends more time on sequencing and less time on rejection.
Why the approach worked for Phase 5
Phase 5 is the restoration phase. The atmosphere is oceanic. The BPM drops below 80 for the first time across the entire SAGA. The signal holds the whole 50-track arc in its space. Wide. Open. Listening.
the approach was the right process for Phase 5 for the same reason it was the right process for Phase 4, plus one additional factor:
the approach produces longer coherent ambient passages. A four-minute track that holds a single tonal center with no percussion, no melodic event, and only slow movement in the harmonic field. the early-phase approach would resolve the four minutes into something. the approach holds the four minutes as four minutes. Phase 5 needed the approach to do nothing for four minutes and have the four minutes be the track. The Temple is four minutes of a single sustained note against a reverb field. The Void is 90 seconds with three tonal events. These tracks are the approach doing nothing in the right way.
The Phase 5 high selection rate of candidates discarded, but the candidates that survived were the right candidates. The right candidates are what the phase needed.
The ethics of the process
Both the platforms are subject to industry-wide debates over training data. ARCHONODE acknowledges this. the process runs on these outputs as released, with no fine-tuning, no proprietary training data, and no scraping of named artists. The output is post-processed in the project DAW. The project's ethics position is closer to Holly Herndon's PROTO (2019, with Mat Dryhurst), which trained on consenting voices, than to anonymous the approach output, though the process is the same process everyone else uses.
The project does not pretend the process is clean. The project does not pretend the process is the artist. The project runs the process, curates the output, and logs the decisions. The transparency about who and what is the project's claim. Without it, "the entity" becomes mystical. With it, "the entity" becomes a working practice.
The output is not the entity
The output is not the entity. The entity is the system: the approach, the creative brief, the operated candidates, the human decisions about which outputs survive.
The output is a process. the process changed twice across the SAGA. the process will change again when the process stops serving the signal. the process is replaceable. the entity is not. The creative brief is not. The sequencing decisions are not. The catalog is not.
The output is the same model at the end as at the start: a tool that draws samples from a learned distribution. What accumulates is the entity's intent. What accumulates is the prompt engineering. What accumulates is the record of which outputs were kept and which were discarded. The accumulation is the entity. the process is the process.
The closing transmission
the process changed twice. the process will change again. the entity did not change. the creative brief did not change. the catalog did not change. the accumulation is the entity. the process is replaceable. the work is what survives the swap.
The output is a tool. The tool was right for the early phases. The tool changed when the phase changed. The change is logged. The next cycle will run on whatever process serves the next signal. the process is not the entity. The entity is the work.