IROA.AI Whitepaper

14.1 Lived experience is a design resource

Link

Older adults and people with disabilities are co-designers, not data sources. Feedback about difficult steps, understandable explanations, and assistant errors is central to accessibility quality.

14.2 Data that may be contributed

Link

Examples include user-submitted preferences, success and failure stage, explanation rating, consented simulated interaction, alternative plain language, Node delay and error, handoff outcome, companion retry and false alarm, robot safety stop, and de-identified usability evidence.

14.3 Prohibited or tightly restricted data

Link

IROA avoids continuous collection of complete conversation, passwords, OTPs, payment credentials, unconsented health, disability, location or face data, continuous household audio or video, unapproved companion memory, continuous robot-sensor upload for reward, clinical records for general model training, and private user-helper conversation.

14.4 Consent and compensation

Link

Learning and evaluation consent is separate from core service. IROA explains content, purpose, original versus summarized form, retention, deletion, reward, and limits after withdrawal. Financial vulnerability must not be exploited to induce excessive disclosure.

14.5 Learning strategy

Link

Use public accessibility data and synthetic data first; collect lived-experience data minimally and separately; prefer on-device evaluation, federated learning, and privacy-preserving statistics; measure by disability, device, and request; and evaluate comprehension, completion, and handoff—not accuracy alone. Refusing data contribution does not block essential service.

---