Multi-model generative system for creating abstract geometric art — seeking the typical form through a Generative Alphabet of primitives.
Each image is a perceptual instrument — a typical form so stable and inevitable that the viewer perceives past and future simultaneously, extracted from time. The system explores geometric primitives as carriers of psychophysical state.
Each generation passes through 10 stages. Every stage is logged to the database.
Three models independently evaluate each image. Consensus score = median of three. Disagreement flag (spread > 3) marks images for human review.
| Voice | Model | Strength |
|---|---|---|
| Structural Analyst | GPT-5 | Prompt-result gap, composition analysis |
| Emotional Intelligence | Claude Opus 4.6 | Deep introspection, state transmission |
| Visual Grounding | Gemini 3.1 Pro | Multimodal analysis, perceptual truth |
| Metric | Weight | What it measures |
|---|---|---|
| State Transmission | 40% | Would the experience HOLD someone who simply existed with the image? |
| Primitive Power | 25% | Is the target primitive the dominant visual force? |
| Perceptual Truth | 20% | Does this feel like being inside a dome? |
| Preset Fidelity | 10% | Does the image follow its assigned preset constraints? |
| Novelty | 5% | Has the agent explored new territory? |
The system maps the space: 8 Experiences × 12 Primitives × 6 Approaches = 616358 cells. Epsilon-greedy selection: 70% explore (least-visited cells), 30% exploit (mutate best cells).
| Stage | Model | Upgrade from V2 |
|---|---|---|
| Image Generation | gpt-image-1.5 | gpt-image-1 → 1.5 (Elo 1264) |
| Scene Plot & Director | GPT-5.4 Pro | gpt-4.1 → gpt-5.4-pro |
| Critic (3 voices) | GPT-5 + Claude 4.6 + Gemini 3.1 | Single gpt-4.1 → 3-model council |
| Reflection | Claude Opus 4.6 | gpt-4.1 → Claude (deeper introspection) |
All prompts use constructive framing: defining what to achieve rather than what to avoid. Every instruction states the target quality positively.
Every stage of every generation is recorded in the pipeline_logs table:
stage name, model used, input/output text (truncated to 5000 chars), duration, token counts,
and the exploration grid cell assignment. This enables post-hoc analysis of which prompt
formulations and cell assignments produce the strongest results.