Every media job in GenCreator is routed by **task modality**, **quality ceiling**, and **credit governance** so a production run never exhausts paid credits without a pre-approved plan.

## The verified model catalog (`TRUTH.md` §3) [#the-verified-model-catalog-truthmd-3]

Through `arcanea-mcp` (powered by `MUAPI_KEY` + the **Gemini TASTE scorer**), the routing engine classifies **226 exact model endpoints** across six modalities:

| Modality | Endpoint count | Primary routing role |
|---|---|---|
| **Text-to-Image (`t2i`)** | `51` | Concept plates, editorial covers, and character base sheets |
| **Image-to-Image (`i2i`)** | `57` | Style transfer, character consistency, and composition refinement |
| **Text-to-Video (`t2v`)** | `43` | Hero motion establishing shots |
| **Image-to-Video (`i2v`)** | `62` | Default motion path — preserves approved still composition at lower cost |
| **Video-to-Video (`v2v`)** | `4` | Restyling and motion transfer |
| **Lipsync (`lipsync`)** | `9` | Dialogue and vocal track alignment |

## Audio & music catalog [#audio--music-catalog]

The audio pipeline operates over **845 verified Suno tracks** (`26,193` plays) across **20 style-voices** (`TRUTH.md` §3), paired with `scripts/music-video/ingest.mjs` for Spotify Canvas (`mv-canvas`), Apple Music motion artwork (`mv-artwork`), and audio-reactive visualizers (`mv-visualizer`).

## Three non-negotiable routing rules [#three-non-negotiable-routing-rules]

1. **Image-to-video before text-to-video.** Lock composition, lighting, and character identity in a still image first, pass it through the **Gemini TASTE scorer**, and only promote passing frames to `i2v`.
2. **Native & NanoBanana first; Higgsfield behind `/hf-plan`.** Per `TRUTH.md` §5a (`[RULED 2026-10-19]`), native image generation (`generate_image` / NanoBanana `nb-image`) runs first. Higgsfield calls are permitted **only after human approval of a written `/hf-plan`** specifying the campaign and credit cap.
3. **Every generated asset writes a provenance sidecar.** Output images and clips must record their prompt hash, model version, provider, seed (when exposed), and critic score alongside the asset.