docs
Platform overview
The three-layer architecture behind every Cadenza generation, and why it exists.
Last updated 2026-07-10
The one idea
Most AI video tools begin by rendering. Cadenza begins by deciding: what the sequence should mean, what each shot contributes, and what must remain consistent. Rendering happens only after those decisions exist.
Three layers
| Layer | Name | What it does |
|---|---|---|
| 1 | Creative reasoning | Turns a brief into a deep, model-agnostic shot plan: the Canonical Shot Object. Story arc, per-shot purpose, camera physics, lighting, pacing. |
| 2 | Per-model encoders | Compresses each shot into the optimal prompt for the specific model it routes to. Depth is tuned per model, never dumped uniformly. |
| 3 | Continuity store | Locks character, style, and lighting references when a sequence starts, and feeds good rendered frames back as references. This is the anti-drift core. |
Adding a new video model touches layer 2 and a routing table. Layer 1 never changes for a new model, which is why model churn does not destabilize the product.
What runs where
- Studio: the web workspace at /studio, for interactive direction.
- REST API: the same engine at /api/v1, for developers.
- MCP server: the same tools for Claude Code, Codex, Cursor, claude.ai, and any MCP client.
- One account and one continuity store back all three. Nothing is a reduced tier.
Safety on the path
Input policy checks run before any model is called; provenance, watermarking, and scanning run after. Publishing requires explicit per-destination confirmation. Details in security and the trust center.