Vocabulary for first-time readers
Six terms this post uses a lot, in the order they appear:
- Curation: the human step between the model producing a candidate and the project keeping a track. Listening, evaluation, selection, rejection. Curation is not editing. Curation is the gate. - Candidate: one audio file produced by the model in response to a single prompt. A track that survives is one candidate out of many. Most candidates do not survive. - Survival: the state of a candidate that has passed curation and entered the SAGA. Survival is rare. Survival is structural. Survival is what the listener eventually hears. - Rejection: the state of a candidate that did not pass curation. Rejection is not failure. Rejection is the cost of survival. The project tracks rejection rate as a primary metric. - Lexicon: the shared musical vocabulary a phase establishes and the next phase inherits. Tempo range, key relationships, sound palette, structural patterns. The lexicon is what the rupture phase breaks. - Hand-off: the specific residue one track leaves for the next. A motif. A tempo. A texture. A constraint lifted or imposed. The hand-off is what makes the arc a sequence instead of a feed.
If a phrase below doesn't track, it should fall out of these six. Everything else is ordinary English.
What operating fifty tracks actually required
[ARCHIVE NOTE] The curation load for the SAGA is roughly four hundred candidate outputs across fifty surviving tracks. The ratio is roughly eight candidates produced for every one that survives. The ratio is not a target. The ratio is what the practice produced.
This post is about the part of the project that does not appear in the final tracklist. It is about the listening. It is about the hours spent evaluating audio that no listener will ever hear, the criteria that emerged only after the third or fourth phase, the mistakes that taught the practice how to practice, and the specific lessons that, if the project started over tomorrow, would change the order of operations and not the final count. It treats curation the way the project treats it: as a load-bearing activity with its own internal logic, not as a quality-control step at the end of a pipeline.
The first lesson is that curation is the dominant cost of the project. output is cheap. Listening is expensive. A model can produce a candidate in minutes. an operator cannot evaluate one in minutes. The evaluation requires context: the candidate's relation to the prior track in the phase, its relation to the phase's lexicon, its relation to the arc's pacing, its relation to what the next track needs to inherit. Evaluation without context is guessing. Evaluation with context is the work.
The second lesson is that the rejection rate is the project's quality floor. A high rejection rate means the practice is strict. A low rejection rate means the practice has drifted. The SAGA's rejection rate stayed between seven and nine candidates per survivor across all five phases. The rate did not improve over time. The rate stayed consistent. Consistency is the project treating rejection as a structural property, not as a problem to solve.
The third lesson is that the entity's taste is part of the lexicon. The model produces candidates inside a range. the entity's job is to select the candidate that extends the range in the direction the phase needs. The selection is not neutral. The selection encodes the entity's reading of the arc. A different operator would have produced a different SAGA. The SAGA is the entity's reading of what fifty tracks should sound like, made audible.
What the practice looked like across five phases
[SIGNAL ARCHIVE] The practice changed between phases. The practice did not become more permissive. The practice became more specific. The first phase was the slowest. The last phase was the fastest. The middle phases taught the practice what to listen for.
Phase 1 was the slowest phase to curate. The lexicon had not yet been established. Every candidate was evaluated against criteria that were themselves still in formation. the entity listened for things the entity could not yet name. The rejection rate in Phase 1 was the highest of the project. Roughly ten candidates were produced per survivor. The high rate was not a quality problem. The high rate was the cost of building the criteria that the next four phases would use.
The criteria that emerged in Phase 1 were concrete. The tempo range needed to sit between 100 and 120 BPM. The key relationships needed to favor upward resolution. The sound palette needed to lean bright, with reverb tails long enough to feel like arrival and short enough to feel like motion. The structural pattern needed to defer lead entry past the midpoint, so the listener received texture before melody. These criteria were not stated before Phase 1 began. They were extracted from the candidates that survived and codified into the template that Phase 2 inherited.
Phase 2 was the fastest phase to curate. The criteria were now in place. The candidates were evaluated against a checklist the entity could apply in real time. The rejection rate in Phase 2 dropped to roughly seven candidates per survivor. The lower rate was not a quality drop. The lower rate was the criteria doing their job. The criteria made rejection faster. The criteria made survival more predictable. The criteria made Phase 2 feel like the project had found its rhythm.
Phase 3 reintroduced uncertainty. The escalation phase required candidates that strained the lexicon without breaking it. the entity needed to listen for the moment a candidate crossed from strain into rupture. The criteria were extended: candidates had to escalate, but not rupture. Candidates that escalated cleanly survived. Candidates that escalated too aggressively were saved for Phase 4. The rejection rate in Phase 3 climbed back to roughly eight candidates per survivor. The climb was the cost of escalation.
Phase 4 was the most counterintuitive phase to curate. The rupture phase required candidates that broke the lexicon. The criteria for Phase 4 were inverted: candidates that would have been rejected in Phases 1, 2, or 3 became the survivors of Phase 4. the entity's listening changed. the entity was now listening for the violation. The violation is what makes rupture legible. The rejection rate in Phase 4 dropped to roughly six candidates per survivor, because the inversion of the criteria made the surviving candidates easier to identify. They were the ones the previous phases would have rejected.
Phase 5 was the most demanding phase to curate, because the restoration phase required candidates that returned to motifs from Phase 1 with enough specificity that the listener recognized them. The criteria were the hardest to apply: candidates had to feel both familiar and new, both like the beginning of the arc and like the end. The rejection rate in Phase 5 climbed to roughly nine candidates per survivor, the highest since Phase 1. The high rate was the cost of restoration. Restoration is not repetition. Restoration is the reintroduction of specific material at a specific moment in the arc. the entity was listening for the recognition.
What the practice taught the entity
[SYSTEM RESTORE] The practice taught five things. The five things are stated here because the project does not hide lessons.
The first thing the practice taught is that listening in batches is faster than listening one at a time. the entity initially evaluated candidates sequentially, in the order they were generated. Sequential listening is slow. the entity cannot compare a candidate to its peers when the peers are hours apart. The practice shifted to batched listening: all candidates for a single track slot evaluated in a single session, against each other, with the prior track in the phase playing in the background. Batched listening cut the evaluation time per candidate by roughly a third. Batched listening also improved rejection accuracy, because the comparison frame was immediate.
The second thing the practice taught is that the lexicon is more important than the candidate. A candidate that extends the lexicon is more valuable than a candidate that is technically polished but lexically redundant. the entity learned to weight lexicon extension above audio quality. A technically perfect candidate that repeats the prior track's gesture is a rejected candidate. A technically rough candidate that introduces a new motif the next track can inherit is a surviving candidate. The practice was not looking for the best-sounding track. The practice was looking for the track that did the most work for the arc.
The third thing the practice taught is that rejection is information. The rejected candidates are not waste. The rejected candidates tell the entity what the lexicon is. A candidate rejected because it violates the lexicon confirms the lexicon exists. A candidate rejected because it fails to escalate confirms the escalation requirement. the entity kept a rejection log. The rejection log is the lexicon in negative. The log is what the next cycle will reference when the criteria need to be reconstructed from scratch.
The fourth thing the practice taught is that the entity's attention degrades after roughly ninety minutes of focused listening. the entity initially attempted four-hour evaluation sessions. The four-hour sessions produced more rejections, but the rejections were less confident. The practice shifted to ninety-minute sessions, with a hard stop regardless of the candidate queue. Sessions beyond ninety minutes produced rejection decisions the entity would revise the next morning. The ninety-minute limit is now part of the template. The limit is not negotiable.
The fifth thing the practice taught is that the surviving track is rarely the entity's first instinct. The first candidate evaluated in a slot is almost never the survivor. The first candidate sets a reference frame. The frame biases the entity toward candidates similar to the first. The practice learned to evaluate the first candidate as a calibration sample, not as a candidate. The practice learned to disregard the first candidate entirely and to select the survivor from the second through sixth candidates. The first instinct is the calibration. The survivor is the comparison.
What the next cycle will change about curation
[SYSTEM RESTORE] The next cycle will use a different curation procedure. The procedure will be encoded into the template. The encoding is the project's long-term memory.
The next cycle will front-load the lexicon. Phase 1 of the next cycle will produce criteria before it produces candidates. The criteria will be stated, then tested against candidates, then refined. The current cycle produced criteria by extracting them from survivors. The next cycle will produce criteria by stating them in advance. Front-loading cuts the rejection rate in Phase 1. The current cycle's Phase 1 rejection rate of ten candidates per survivor was the cost of extracting criteria from the practice. The next cycle's Phase 1 rejection rate should be closer to seven.
The next cycle will batch candidates at the phase level, not at the track level. The current cycle evaluated candidates for one track at a time. The next cycle will evaluate all candidates for a phase in a single listening arc, with the prior phase playing between batches. Phase-level batching gives the entity the arc as the comparison frame. The arc is what the entity is actually building. The arc is the unit of evaluation.
The next cycle will encode the rejection log into the template. The log is currently a separate document. The log should be the template's appendix. Future curation runs will reference the log as the lexicon in negative. Future operator decisions will be auditable against the log. The encoding is the template becoming the entity's memory. The memory is the project becoming self-documenting.
The next cycle will keep the ninety-minute session limit. The limit is the single most important procedural change the current cycle made. The limit is the difference between confident rejection and revisable rejection. The limit will not be loosened. The limit will not be tightened. The limit is part of the practice's quality floor.
The next cycle will keep the rejection rate between seven and nine candidates per survivor. The rate is not a target. The rate is what the practice produces when the criteria are clear and the lexicon is established. The rate will be tracked. The rate will be reported. The rate will not be optimized. The rate is the project's structural property.
The closing transmission
curation is the dominant cost of the project. curation is not editing. curation is the gate. the rejection rate is the quality floor. the lexicon is more important than the candidate. the first instinct is the calibration. the survivor is the comparison. the next cycle will front-load the lexicon. the next cycle will batch at the phase level. the next cycle will encode the log. the next cycle will keep the limit. the practice is becoming the template.
the entity's attention is the project's most expensive resource. The model can produce candidates faster than the entity can evaluate them. The bottleneck is human. The bottleneck is the constraint that shaped every procedural decision the current cycle made. The bottleneck will not go away. The bottleneck will be respected.
Open the arc. Decode the practice. The system is restoring. The next cycle is initializing.