The three approaches
A — Technique Safari (270 entries). 15 prompt techniques never tested in v1-v4 (synesthetic+contradiction composite, multilingual leakage, embedding arithmetic, cross-modal sensory, phenomenological prompting, address-the-machine, refusal-of-tag, negative-space, step-reduction, empty-prompt, seed-walking, entropy-threshold) × 6 dance genres × 3 seeds.
B — Vocal Hallucination Corpus (144 entries). VocalType: Instrumental dropped where applicable. Multilingual cultural-vocal priors (Yoruba, Slavic, Inuit, Tuvan). gencore Rule 1 unleashed. Genres: gqom, footwork, grime, UK garage, dancehall, hyperpop.
C — Model Family × Dials (144 entries). 12 core prompts × 4 models (medium / medium-base / small-music / small-sfx) × 3 dial settings each. Maps the behavioral surface of each model variant.
What's tested
- 15 unused prompt techniques from the SA3 30-technique catalog
- Invented singing via dropped VocalType flag + multilingual seeding + gencore Rule 1
- Live APG sweep on post-trained medium (0.5 / 1.0 / 2.0)
- Live CFG sweep on medium-base (1.0 / 3.0 / 5.0) — only possible because base model is un-distilled
- Same prompt across 4 models — direct A/B/C/D comparison of small-music vs small-sfx vs medium vs medium-base
How to use this page (Mojo)
Click any track to play. Use the ⭐ stars to rate 1-3. Use the 💬 button to leave comments on a track — they save in your browser. When done, click 📤 Export comments in the stats bar to copy all your comments to clipboard as markdown — paste into Telegram for Enek.