IROA.AI Whitepaper

11.1 Technology layers

Link

IROA combines speech and multimodal interaction, task planning, official tool integration, policy enforcement, isolated browser execution, outcome verification, accessibility adaptation, human handoff, and Node attestation. No single model is trusted to enforce every boundary.

11.2 Model selection

Link

Models are selected by task, language, latency, device, cost, privacy, and error profile. Small local models can handle wake words and simple classification; stronger hosted or managed models may plan complex work after minimum-data filtering. Critical approval, policy, and outcome checks use independent rules and service responses.

11.3 External-service connections

Link

Official APIs and standards come first, tool contracts validate input and output, and screen execution is isolated and monitored. Each integration publishes supported actions, permission requirements, failure behavior, and evidence quality.

11.4 Starting with limited capital

Link

IROA begins with existing models, open standards, regulated payments, a small catalog of high-value requests, managed cloud isolation, and institution-led pilots. It does not begin by training a foundation model, replacing hospital records, manufacturing a general robot, or building an unrestricted automation platform.

11.5 Research and development scope

Link

Priority research includes disability-specific interaction, plain-language generation, robust outcome verification, prompt-injection resistance, capability security, remote attestation, deletion proof, on-device companion and wearable models, robot stop policy, accessibility evaluation, and privacy-preserving learning. Research claims remain separate from deployed capability.

---