Pipeline History (v1–v17)
This is the archived pipeline description. Current version →
1
Snapshot — the system collects real-time data from 14+ sources (Wikipedia, weather, earthquakes, moon phase, poetry, music, art galleries, ocean tides...) and interprets it through an LLM into a visual ontology: entities, forms, materials, transformations, mood, and connections.
2
Scene Director — an LLM freely composes a cinematic scene based on the ontology, emotional intent, accumulated visual memory, user feedback, and the current hypothesis being tested.
3
Image Generation — the scene description is rendered into an image. Two images per cycle: one tests a new hypothesis, the other is a control — an A/B test.
4
Critic — evaluates images against composition rules, ontological richness, emotional contract, and transcendence. Compares intent vs result using the actual scene prompt.
5
Reflection — the system analyzes what worked and what failed using score deltas. Proposes the next hypothesis as a testable experiment.
6
Meta-reflection — analyzes the reflection process itself. Evaluates whether its own patches improved how the system thinks. Traces the full causal chain: patch applied → reflection changed → scores affected.