Mossferaitu · v4 — Machine-Side Prompting

v4 — Machine-side prompting at the conditioning boundary. Generated from the SA3 hyperresearch catalog (30 techniques in three surface families: conditioning-side, sampling-side, composition-side). This corpus tests the report's claims empirically. APG sweep, init_noise_level sweeps, geometric-practice prompts (Eneki), and the 50-prompt arc corpus from PART A.
The six practices (Eneki's geometric prompting note)
  • Single-axis movement. Move along one dimension at a time. Stacking modifiers forces void-crossing.
  • Describe paths, not points. Trajectories through concept-space scan more cleanly than static descriptions.
  • Verbalized sampling. Generate variations along explicit axes (metallic / organic / distant / intimate / abstract). Map the local manifold before committing.
  • Modular decomposition. Combine simple features stepwise; complex composite prompts may not resolve cleanly.
  • Follow the bidirectional geometry. When two concepts compose well, they share latent proximity. Build a mental map.
  • Don't over-interpret single outputs. One generation = one point on the manifold = blind man's fragment. Generate multiples and cluster.
What's being tested

Pre-dawn (25 prompts). Synesthetic chains, multilingual leakage (Finnish / Japanese / German), contradiction stacking, verbose-prose prompting, diegetic world-prompts, cross-modal sensory transfer, embedding arithmetic via plus-strings, negative-space prompting. The pre-dawn arc target: sub-bass + irregular + descending = max temporal dilation.

Dawn-cross (10 prompts). Transitional moves: sub-bass ascend, contradiction-triangulation, dissolution-into-pulse, multilingual ascent, motif emergence.

Post-dawn (15 prompts). Verbose-techno, A4-A5 register (PTIU peak), gencore Rule 2 (incoherent genre stacks), forced glossolalia, embedding arithmetic, dub-techno, jungle, half-step dubstep, EBM, motivic-recurrence. Post-dawn target: mid-register + locked grid + ascending = compression, with familiarity-illusion anchors.

Experiments (~25 prompts). APG sweeps (apg_scale ∈ {0.5, 1.0, 1.5, 2.0, 2.5} at fixed seed), init_noise_level sweeps (0.4 / 0.6 / 0.8 / 0.9 over v3 source), Eneki's six geometric practices (single-axis, path-prompts, verbalized sampling, geometric proximity tests), manifesto-as-constraint.

Each track is rated 1–3 ⭐; ratings persist to ratings.jsonl for next-corpus design.

—