# AETHER-Ad Genesis — Architecture (v0.2)

## Layered pipeline

```
Layer 0  Meta Framework ──────────────── schema/meta_framework.json
Layer 1  Product Genome ──────────────── schema/product_genome.json   + core/genome.py
Layer 2  Context Matrix ──────────────── schema/context_matrix.json   + core/context.py
Layer 3  Collision Rules (15) ───────── schema/collision_rules.json  + core/collision.py
Layer 4  Narrative Scaffold ─────────── schema/narrative_scaffold.json + core/narrative.py
Layer 5  Wow Filter ─────────────────── schema/wow_filter.json        + core/wow_filter.py
         Genesis Engine (orchestrator) ─ core/engine.py
```

## 5-stage pipeline (Young 1940, via AETHER)

1. **Corpus Ingestion** — pull atoms + tension + persona + forbidden zones.
2. **Encoding** — `blending/spaces.py` builds the 4-space diagram (I1, I2, generic, blend).
3. **Incubation** — sample N collision rule combinations, call LLM, produce concepts.
4. **Emergence** — render each concept via Pixar Story Spine (15s/30s).
5. **Filtering** — Wow Filter (5 axes + risk + gating) → top-k.

## SLAI feedback loop

Approved top-k seeds → `engine.slai_feedback()` → proposed new collision rule draft → human review → added to `collision_rules.json`. This is the advertising-domain implementation of AETHER's Self-Learning-AI principle.

## Backends

- `HFInferenceBackend` (default for Space) — `huggingface_hub.InferenceClient`.
- `FireworksKimiBackend` — Fireworks Kimi-K2P5 (OpenAI-compatible).
- `DarwinOpusBackend` — local Darwin-27B-Opus endpoint on H100 (v0.3+).

All backends implement the same `LLMBackend` contract (`complete` + judge helpers).