v12 stopped asking SA3 for whole tracks. v1–v11 cooked songs end-to-end; v8 proved the gravity wells lock the model into the training centroid; the morphable space lives between them. v12 cooks ingredients — single-role sounds with no compositional claim — and lets an assembler own the spatial grammar.
Three components: cooker generates per-role ingredients via SA3 prompts (BPM-locked, key-locked at cooking time), library catalogs them with metadata, assembler reads a section template (gradient-arc or plateau-grammar) and an ingredient pick-list, builds per-role timelines with mute automation, applies filter automation, Gross Beat stutter for hook-substrate roles, sidechain compression on bass keyed by kick, then mixes via FFmpeg and masters to -8 LUFS.
Architectural source-of-truth: the dj-tool-track-anatomy-ccda3d hyperresearch run (10k-word adversarially-audited report on what makes exceptional 2024-26 DJ-tool tracks). Every design choice traces back to a specific corpus claim.