The Liquid Canvas: A Reference Architecture for Governed Generative User Interfaces in Regulated Enterprises

Résumé fourni par la source

Interfaces that compose themselves for each person are technically feasible, yet in regulated sectors — banking, insurance, healthcare, government — they have seen almost no production adoption. The barrier is not capability but provability: an institution cannot demonstrate, after the fact, what an adaptively generated interface displayed, nor guarantee, in advance, what it will never do. This paper presents the Liquid Canvas, a reference architecture for governed generative user interfaces — an agent-composed form of server-driven UI, in which the client still renders a server-issued specification but a governed agent, not imperative server code, decides what that specification contains. Its central move is to remove rendering authority from the model. The agent emits a declarative layout contract — a validated selection over a bounded library of vetted components, bound to sources of record — rather than markup, data, or copy. The architecture comprises seven components — a component catalogue, a contract schema and validator, brand tokens, a statechart, a risk-banded approved-copy store, read-only data access over the Model Context Protocol, and a version-and-audit ledger — governed by one discipline: the system re-plans against observed state under fixed policy. It adapts; it never learns. We describe the composition pipeline, a tiered escalation model that invokes agency only when simpler mechanisms fail, and the governance properties — determinism, replayability, human-in-the-loop gating, and provenance — that follow. We then map these properties to ISO/IEC 42001, ISO/IEC 27001, SOC 2, India's DPDP Act and Rules, and the EU GDPR, distinguishing what the architecture evidences from what the deploying organisation must still supply, and illustrate the approach with a governed retail-banking demonstrator.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.