ECHOFRAME — Cinematic Intent Lock

What it does

ECHOFRAME is a governed pre-generation control layer that converts cinematic intent into executable crew contracts, blocks mismatched requests before provider dispatch, verifies generated pixels, and preserves every decision through the official ClickHouse MCP.

Three AI film crews may all return plausible images while still failing the edit. ECHOFRAME first turns the director's needs into distinct executable contracts: spatial coverage, rescue action, and safe handoff. A human approval boundary keeps provider dispatch at zero until those contracts are accepted. Requests that violate an approved contract are rejected before they reach the provider. Generated frames are then checked at pixel level; a plausible first Crew C frame was rejected because it repeated Crew B's rescue action instead of delivering the required emotional payoff. Human correction produced the verified safe-handoff frame.

The failed output was not deleted or rewritten. The rejection, correction, verification, and final ACCEPTED_WITH_NONCRITICAL_PROP_DEVIATION adjudication were appended to the ClickHouse audit history.

How we built it

• Google ADK and Gemini propose the three cinematic contracts.
• A fail-closed Cinematic Intent Lock requires named human approval before provider dispatch.
• Google image generation creates only approved coverage requests.
• Gemini performs pixel-level contract verification on the generated images.
• The official mcp-clickhouse==0.4.1 server records and reads the append-only decision history through run_query over stdio.
• Python contract tests exercise proposal recovery, payload mapping, reference chaining, dispatch authorization, and zero-retry behavior.

Evidence and factual boundary

The demo video is a verified-state replay, not a recording of one continuous live session. Generation, Gemini verification, and ClickHouse MCP evidence are real execution results. The interface replays those verified states and does not claim a continuously running cloud product.

The hosted app is an earlier, clearly labelled simulated control-room surface; its rail scenario is not part of the Cinematic Intent Lock generation set or the final submitted film evidence.

Actual call ledger:

• ADK: 1
• Images: 6 attempted / 5 completed / 1 rate-limited
• Pixel checks: 2
• Automatic retries: 0
• Conservative billing cap: $0.191

Built with

Google ADK, Gemini, Google image generation, Vertex AI, Python, Model Context Protocol, official mcp-clickhouse 0.4.1, ClickHouse, TypeScript, React, and vinext.

Built With

• context
• google
• model
• vertex