A rights-clean composer-model scoring house.
The model sketches; the composer ships. BOUNCES turns Simon Taylor's artist-owned source bounces into seed-conditioned cue candidates, grades them by ear, and ships winners with provenance receipts: source lineage, dataset hash, taste note, license chain, and musical inheritance evidence.
Why this is the project
Fine-tuning MusicGen is commodity work. A known composer using only his own catalogue, selecting seeds by brief, generating cues in his idiom, judging by ear, and issuing receipts is not commodity work.
The training and seed pool is constrained to Simon-owned source bounces. That constraint is the moat: no unclear dataset, no borrowed style, no other artists' material.
Every output is graded by ear: P for portfolio candidate, T for texture, B for broken. The taste log is not admin; it is the first training set for a future Simon's-ear reranker.
Each winning cue carries source lineage, dataset hash, generation path, ownership terms, and musical-inheritance notes. Post-lawsuit music AI needs proof, not vibes.
Workflow as artifact
This is the agentic workflow case study: a practical pipeline that a creative team, AI studio, sync house, or brand lab can understand in one pass. The same discipline connects SALIX's LangGraph/pgvector world logic, the GSS source-audit pipeline, and the receipted BOUNCES pilot.
Supervisor sends text, temp, tone, duration, and use case.
CLAP-style seed retrieval finds candidate source bounces from the owned bank.
MusicGen-Melody sketches variants from selected seeds.
Simon grades by ear: phrase, harmonic intent, idiom, production, seam.
Human finish pass turns a sketch into a shippable cue.
Winning cues ship with provenance and musical inheritance notes.
Definition of shipped: a supervisor Simon did not personally brief plays a cue for a client, and the cue carries a receipt. Everything before that is rehearsal.
Batch 2 evidence
Batch 2 has twelve generated cues. The first confirmed P grades are below: Embrace the Serpent and Alone Manhattan Place. Both are public-safe generated audio loops copied into this portfolio build with receipt manifests.
Melnyk-like avant-piano behavior; white-noise spectral thickening around 12 seconds; perfect loop seam. The MIDI transcription shows a monophonic line with spectral development rather than harmonic padding.
Haunting, odd melody take that turns more complex; a clean loop with wide-register leaps and a stable G#/F pitch-class centre.
receipt: batch2-03 / Embrace the Serpent status: portfolio_candidate generated_audio_sha256: 47451df6f06cba87a0b45203d3288d956aba28ea7d42df9ad0a379d670d4ac47 source_lineage: EMBRACE_THE_SERPENT_mix1 - artist-owned source material dataset: 159 eligible bounces / 580 prompt-audio pairs taste_grade: P - Melnyk avant-piano; spectral thickening; perfect loop musical_inheritance: 44 MIDI notes, B2-A#5 range, C and F pitch-class centres human_decision: model sketch accepted as portfolio candidate; composer finish still required for sync use
WebAudio / spatial-DSP micro demo
This closes the spatial-audio proof gap without inventing a fake app. The demo uses a real generated BOUNCES cue and a browser WebAudio panner. It is deliberately small: play the loop, move the source left/right and front/back, and hear the cue behave as an object rather than a flat file.
The recruiter signal is not "DSP engineer." It is: composer-builder can prototype spatial behavior directly in the browser, then explain the path to Wwise/Unity.
Adjacent proof surfaces
| Gap | Status on this page | Next build step |
|---|---|---|
| Agentic creative-workflow case study | Addressed - BOUNCES workflow from brief to receipt. | Build CLAP seed retrieval and receipt generator as runnable local tools. |
| SALIX-derived ComfyUI / ControlNet / LoRA workflow sheet | Framed - keep as a workflow artifact, not a loose case-study essay. | Publish Style Clock as a separate executable sheet once visual outputs are selected. |
| WebAudio or Unity spatial-DSP micro demo | Embedded - browser spatial panner using an actual BOUNCES cue. | Add Wwise/Unity mapping notes after the SALIX capture pass. |
Immediate next steps
Finish Batch 2 grades, build seed retrieval over the 159 bounces, generate human-readable receipts for every P cue, and chain :60/:90 continuations only after receipts exist.
Run one friendly sync-supervisor brief end to end: real temp, real deadline, three candidate cues, receipts visible, and a human finish pass. That is the line between demo and product.