{"execution":"POST /api/demo/full returns live, fail-closed model receipts","models":[{"name":"gemini-3.5-flash","purpose":"quote-grounded synthetic artifact extraction (fixture image and supplied advisory text)","docs":"https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-5-flash"},{"name":"gemini-embedding-001","purpose":"semantic evidence routing without authority decisions","docs":"https://docs.cloud.google.com/vertex-ai/generative-ai/docs/embeddings/get-text-embeddings"},{"name":"gemma-4-26b-a4b-it","purpose":"facility reply triage through the Gemini API; flags mismatches for human review, never decides","docs":"https://ai.google.dev/gemma"},{"name":"gemini-3.5-flash (managed Agent Runtime)","purpose":"ADK role agents propose a typed command; a post-model schema guard decides what is allowed","docs":"https://google.github.io/adk-docs/"}],"inline_guardrails":"Model Armor sanitizeUserPrompt on supplied advisory text and facility notes; blocked text returns HTTP 422 with the receipt","memory":"Vertex AI Agent Engine Memory Bank recalls prior facility behaviour on intake and remembers outcomes on closure","replay_policy":"recorded outputs are test-only; deployed full workflows do not silently substitute them"}