CineTrust watsonx: Explainable Film Risk Triage

Inspiration

Film teams need to catch production risks before cameras roll, but automated flags can easily be mistaken for legal clearance. CineTrust brings explainable screenplay triage earlier while keeping final authority with qualified people.

What it does

CineTrust turns a screenplay scene into inspectable workflows for:

entity and trademark-risk triage; content-safety and age-suitability review; IBM Granite dialogue-alignment integration; provider-health and MCP tool inspection; and deterministic screenplay-entity findings produced by the IBM Bob-authored rules engine.

Every result preserves the boundary that CineTrust is decision support—not legal advice, copyright clearance, certification, or an official rating.

How we built it

FastAPI and Pydantic for the API and structured evidence contracts Google GenAI SDK and a Vertex AI path for Gemini 3.7 Flash Official ibm-watsonx-ai SDK with an IBM Granite ModelInference path A three-tool IBM MCP server for IP triage, guardrails, and dialogue alignment An authenticated IBM Bob development session that added the deterministic screenplay-entity risk-taxonomy service and POST /api/v1/trust/audit/entities Explicit demo, live, live-unavailable, and live-error states with no silent fallback A self-contained judge interface and embedded cinematic-evidence screening room

The public judge workflow now runs live Gemini 3.7 Flash through Vertex AI ADC followed by the deterministic IBM Bob-built risk engine. The separate watsonx/Granite SDK path remains visibly unconnected because no IBM Cloud inference credentials are exposed; no live watsonx claim is made.

IBM Bob development evidence

IBM Bob was used in an authenticated participant-owned development session. Bob added the deterministic entity-risk engine, API route, and focused tests. The dated session record, screenshot, code diff, and independent verification are public in the repository under docs/evidence/IBM_BOB_DEVELOPMENT_SESSION_2026-08-30.md.

Challenges

The central challenge was preventing a governance product from overstating its authority. We removed unsupported compliance claims, fail closed when a live provider is unavailable, expose provider state clearly, escape rendered report content, and require human legal review.

Accomplishments

Four coherent judge-facing governance workflows Explainable, input-dependent scene and entity triage Official watsonx SDK and MCP implementation paths Authentic IBM Bob-assisted development evidence Strict provider-state and failure propagation 35 passing backend tests plus the cinematic-evidence verification test (36 total) Public MIT-licensed repository and verified production deployment Public sub-three-minute functional walkthrough with English captions

Supplemental cinematic evidence

When the Lights Came Back is a human-reviewed 56-second, seven-shot romantic short produced as fixed pre-generated evidence with Gemini 3 Pro Image, Veo 3.1 Fast, and a Lyria 3 Pro score. It demonstrates a continuity-sensitive film asset CineTrust can evaluate. It is not generated on demand by the hosted application and does not establish live watsonx inference or automated legal clearance.

Judge it

Open the live application and compare two screenplay presets. Inspect the explained entity and risk findings. Try Guardrails, Granite Dialogue Alignment, and MCP/System Health. Review the IBM Bob evidence in the public repository. Watch the primary application walkthrough, then the supplemental continuity film.

Built With

• fastapi
• gemini-3.7-flash
• google-genai-sdk
• ibm-bob
• ibm-granite
• ibm-watsonx-ai
• model-context-protocol
• pytest
• vercel
• vertex-ai