Zoetrope: A Deadline-Aware Architecture for Continuous Multimodal Perception, Reasoning, and Authorized Action

A live visual assistant must do more than describe successive images: it must maintain a coherent account of an evolving situation, allocate reasoning effort before relevant opportunities expire, and distinguish an authorized action from a plausible but outdated suggestion. We propose Zoetrope, a model-agnostic architecture for continuous multimodal perception, asynchronous reasoning, and bounded intervention. Its starting configuration samples every other frame of a video stream. Rather than equating this sampling frequency with the frequency or latency of complete reasoning, Zoetrope separates observation ingestion, temporal state maintenance, reasoning admission, and action execution. A versioned evidence store connects claims to timestamped observations; a deadline-aware scheduler allocates deeper inference; and an independent execution gate revalidates permissions and action-critical evidence before an intervention. The design supports both responsive questions and standing requests that require proactive notice or action, while perception continues during response generation. We formalize the distinction between throughput and freshness, specify bounded-memory and overload behavior, and develop a causal evaluation protocol that measures deadline-conditioned correctness, stale-state errors, interruption burden, and resource use. The central hypothesis is that explicit evidence validity and selective reasoning improve useful real-time performance over isolated frame analysis and asynchronous systems without freshness checks. This manuscript presents an architecture, analytical constraints, and an experimental plan; it reports no implementation measurements or benchmark results.