From 81ae22d01a695750067232c80befb1676dc2494d Mon Sep 17 00:00:00 2001 From: Meta Vens Date: Sun, 31 May 2026 09:03:02 +0100 Subject: [PATCH] Record paired song curriculum pass --- state/evidence.jsonl | 1 + state/scratch.md | 10 ++++++++++ 2 files changed, 11 insertions(+) diff --git a/state/evidence.jsonl b/state/evidence.jsonl index f43187c..1479ce1 100644 --- a/state/evidence.jsonl +++ b/state/evidence.jsonl @@ -254,3 +254,4 @@ {"date":"2026-05-31","area":"song-scene-texture-owner","lesson":"The album-swarm patches exposed a bad scene-noise owner: agents modeled air/dust/recording condition with full-duration broad `wave=noise` voices, producing static white hiss that sounds like broken subtractive synthesis. The DSL now has `texture`/`noise_texture` sugar with role defaults for dust, air, codec, room, tape/vinyl; it lowers into a gated noise voice with bandpass/highpass/lowpass/tremolo shaping and an owned gate curve or pattern. Future song prompts require texture or an equivalently gated/band-shaped explicit voice for background noise.","evidence":"src/AquaSynth.Core/PatchScript.cs; tests/AquaSynth.Dsl.Tests/PatchScriptTests.cs; tools/run-song-snippet-trial-swarm.ps1; docs/audio-doctrine.md"} {"date":"2026-05-31","area":"song-source-corpus","lesson":"The song-snippet swarm source corpus now defaults one folder up to `D:\\Music\\Reinier\\Heyoka` and recursively discovers supported audio files, so agents sample across the Heyoka `Gate Code`, `Mandelbass`, and `Marklar` folders instead of only one album folder.","evidence":"tools/run-song-snippet-trial-swarm.ps1; local corpus discovery found eight supported Heyoka tracks under the parent folder."} {"date":"2026-05-31","area":"song-curriculum-dataset-distillation","lesson":"The song trial loop trains the dataset, not a single patch. Each agent now gets paired song challenges and must write one patch per target; every patch, hypothesis, metric, challenge artifact, evaluator fact, and distillation document becomes curriculum evidence in CultCache/Qdrant. The post-pass distiller writes pruned `.cc` stores with reusable scene roles, transfer rules, failure patterns, kept trials, dropped noise, and aggregate metrics, and retrieval links those distillation chunks back to kept scored trial IDs so future zero-shot production agents can actually receive the distilled context.","evidence":"tools/run-song-snippet-trial-swarm.ps1; tools/IpaTrialWorker/Program.cs; src/AquaSynth.Core/IpaTrialResults.cs; state/scratch.md"} +{"date":"2026-05-31","area":"song-curriculum-paired-pass","lesson":"A paired song-curriculum pass completed with five agents, two targets per agent, one patch per target, ten scored trials, ten target evidence bundles, sixty Qdrant chunks, and a distilled `.cc` store. Metrics were all `pressure`: best log-mel cosine `0.502128`, mean `0.377955`, worst `0.27853`, mean audio score `0.339432`. Distillation kept all ten and identified clocked dust/hat texture plus rubber/sub bass as reusable roles, while preserving the shaped-noise owner requirement as a transfer rule rather than letting static white hiss back into the curriculum.","evidence":"artifacts/parity/song-snippet-swarms/20260531T074935298/song-trial-results.cc; artifacts/parity/song-snippet-swarms/20260531T074935298/pass-001/distillation-report.md; artifacts/parity/song-snippet-swarms/20260531T074935298/pass-001/vector-index-after.md; artifacts/parity/song-snippet-swarms/20260531T074935298/pass-001/semantic-search-after.md"} diff --git a/state/scratch.md b/state/scratch.md index ce32472..34a1b07 100644 --- a/state/scratch.md +++ b/state/scratch.md @@ -1775,3 +1775,13 @@ Next likely slice: - Semantic retrieval now indexes distillation documents as chunks linked to kept scored trial IDs, so the search merge can project distilled lessons into the next agent context instead of dropping them as orphan curriculum prose. +- One paired-curriculum pass completed at + `artifacts/parity/song-snippet-swarms/20260531T074935298`: five agents, two + targets each, ten scored patches, ten challenge evidence bundles, sixty + Qdrant chunks, and a distilled `.cc` store. All ten verdicts were `pressure`; + best log-mel cosine was `0.502128`, mean `0.377955`, worst `0.27853`, mean + audio score `0.339432`. +- Distillation kept all ten trials and summarized reusable roles as clocked + dust/hat texture plus rubber/sub bass. That is useful but still not victory: + the scene-noise lesson remains that background/recording color must be a + shaped role owner, not static hiss.