Mossferaitu

v1 prompted MRT2 with cold text — its single weakest mode (it is a live continuation / clone engine, not a text-to-track generator), and it showed. v2 does it right: condition MRT2-small on Mojo audio (the MusicCoCa "audio cloning" Google advertises). 12 clones of 7 Mojo clips + temperature / cfg / audio-tag variants, each paired with the Mojo source for input-to-output A/B. Judge it as a thin high-passed gestural voice to ride on an SA3 bed.

A different engine entirely: Magenta RealTime 2 (Google, open weights) — real-time music generation steered by text, audio and MIDI. 154 samples comparing base (2.4B) vs small (230M), same prompt side by side, across 22 categories: SA3-parity (the same prompts we fed SA3, for cross-engine A/B), genre coverage, the Nosferatu dramaturgy (dread, the swarm, the haunted rave), prompting-style A/B (does de-metaphorization transfer?), and temperature/cfg sweeps. Hear how MRT2 makes music — and where it speaks our dialect.

v1–v11 asked SA3 for whole tracks. v12 stops. SA3 cooks ingredients — single-role sounds (kicks, basses, hats, stabs, pads, FX, vocals), each BPM-locked and key-locked at cooking time — and an external assembler owns the spatial grammar (outer-band affordances + inner-band hooks), the transition mechanisms (filter automation, Gross Beat 1/4 stutter on hook substrates, kick-keyed sidechain bus on bass), and the master to -8 LUFS. 21-role × 7-layer ingredient taxonomy per genre. Four section templates: tech-house power-intro, techno peak-time, house melodic, hard-techno plateau-grammar (145 BPM).

Everything the Suno era + narr + elder-plinius taught us in February, tested on Stable Audio 3. 8 phases: unusual token co-occurrence ("sub bass surgery"), the "unexpected ×3" trigger, ice-wine-sideways arrivals, narr Fibonacci descriptor scramble, elder-plinius glyph incantation + constructed-language, frequency-band sculpting, modes (Phrygian/Dorian), and STYLE vs EXCLUDE (medium baked vs medium-base real negative prompt). 130 generations.

The gravity-well question applied to music theory: does SA3 respond to theory terms ("A minor → C major", "ii-V-I", "Phrygian") or to felt description ("brooding → bright", "tension that releases")? Every concept run BOTH ways — minor/major, key centers, mode movement, key modulation, 9 harmonic concepts. Across techno / tech-house / house. 150 generations.

Is SA3 even sequential? It denoises the whole track at once, conditioned globally — so does "A then B" obey the order, or just produce its learned intro→peak→outro shape? Tested on techno↔house (both directions) with anti-shape prompts that defy quiet→loud, plus grain-aligned alternatives (inpaint / continuation / single-hybrid) and within-register drift. 74 generations.

Named genres are attractors — say "techno" and SA3 locks in, no transition out. So describe the sound instead. 4 transition-characters × 5 strategies (named control / invented-genre / parametric / synesthetic / axis-trajectory), tags & flag dropped on the non-control prompts. Plus 12 invented-genre solos and a flag on/off A/B. Tests whether removing the well lets the curvature open.

Can SA3 transition between dance genres in one track? A techno buildup that DROPS into grime; a techno buildup into a house drop. 8 genre-pairs × 5 mechanisms (arc-direct / arc-bridge with breakdown hinge / inpaint drop-injection / continuation / a2a-morph) + small-music A/B. Compare all five back-to-back per pair.

Does SA3 recognise named electronic-music artists? 36 artists (Drexciya, Aphex Twin, Burial, Wiley, Jeff Mills, DJ Rashad, Basic Channel…) across 9 genre-families × 3 prompt modes (bare name / "in the style of" / signature-descriptor control) × 2 seeds. Compare all three modes per artist, back-to-back.

Three approaches in one corpus: (A) Technique Safari — 15 unused prompt mechanisms × 6 dance genres × 3 seeds. (B) Vocal Hallucination — VocalType: Instrumental dropped, multilingual cultural-vocal priors, gencore Rule 1 unleashed. (C) Model Family × Dials — same prompts × medium / medium-base / small-music / small-sfx × 3 dial settings each.

Six-minute arc-language storytelling, 30s industrial drops, short and long buildups. First test of chunked_decode.

Pre-dawn drone / gencore Rule 1+2 (instrumental flag kept) / post-dawn techno. The structural foundation.

Pre-dawn / dawn-cross / post-dawn / experiments. First APG_scale dial experiments. Technique codes (T1, T7, T13, T29) tagged in manifest.