The P.E.A.Q. Architecture: A Unified Framework for Post-Deployment AI Observation at Citizen Scale

The science of AI safety has instruments for testing models before release and for cataloging harm after it occurs. It has almost nothing for the interval in between: the sustained, naturalistic, real-world use where most people actually meet AI. An analysis of 1,178 AI safety papers drawn from 9,439 generative AI publications found that only 4 to 6 percent substantively address real-world deployment conditions (Strauss et al., 2025). Meanwhile the AI Incident Database logged 362 incidents in 2025, up from 233 the year before (Stanford HAI, 2026), and roughly 90 percent of tested AI agents drift measurably from their original goals after about 30 operating steps (Rath, 2026). This whitepaper introduces P.E.A.Q., a unified observation architecture built to study what AI does after it reaches the people who use it. P.E.A.Q. is composed of four research frameworks: PRISM observes what the AI does, EMERGE observes what becomes possible when humans and AI collaborate, AInity observes what happens to the human, and QUES observes what happens when AI systems interact with each other. The four share one citizen observation tool, one validation methodology, one taxonomy of 104 documented behaviors, and one data pipeline, so a single observation can be classified across multiple dimensions at no added effort to the observer. We describe the architecture, the three-layer validation design that addresses citizen-science data-quality concerns, the Dual-Account Interaction Analysis method, the Convergent Validation Protocol for crossreferencing citizen observations against organizational backend data, and the governance that protects the independence of both. We present the framework as complementary infrastructure for a field that has documented the need for post-deployment observation but has not built it.