When an AI agent must not emit a prohibited practice or unlabeled synthetic output, the check belongs on the prompt and the output, now. That is ComplyEdge. This is not legal advice.
OpenAI’s Article 5 Help Center tells deployers they must not use the models for prohibited practices. It names no runtime vendor. ComplyEdge is the runtime deny: Article 5 and Article 50, article-cited, on this I/O.
Classifiers named eu-ai-act-* score the system. ComplyEdge denies this prompt or output. Article 50 here is unlabeled or deceptive use, not C2PA watermarking.
Bedrock, Azure, and Vertex safety filters stay in place. ComplyEdge sits on top as article-cited deny, not a replacement for Model Armor.
GOPAL is an OPA library you run yourself. ComplyEdge is per-request deny, citation, trust page, and MCP.
CI: uses: complyedge/trustlint-action@v1
npx -y @complyedge/mcp
pip install complyedge
pip install trustlint
pip install 'complyedge[mcp]'Hosted MCP: https://mcp.complyedge.io/mcp. Offline MCP is TrustLint regex, not hosted OPA.