CRP: AI 시스템의 멀티스레드 인지 인식을 위한 개방형 프로토콜

Introducing the Cognitive Resolution Protocol — an open standard for how AI systems should actually see. There's a moment every experienced worker recognizes. You glance at a screen full of numbers and something feels wrong before you can say what. Your eyes haven't finished scanning the columns, but a signal has already fired — a blend of pattern recognition, emotional context, and accumulated expertise that says pay attention. Current AI systems don't do this. They process. They receive input, run it through a model, and produce output. One input, one model, one answer. That's not perception — it's a function call. Today we're publishing a research paper and protocol specification for something we believe is fundamentally missing from AI architecture: multi-threaded cognitive perception. When a browser automation agent looks at your screen, it sees pixels. It runs those pixels through a vision model and gets a text description. That description goes to a language model that decides what to do. One thread. One interpretation. No nuance. A human looking at the same screen processes it simultaneously across multiple cognitive channels: task recognition, anomaly detection, emotional response, privacy awareness, temporal urgency, and memory. And those channels aren't independent — they talk to each other. The anomaly detector asks memory for historical comparison. The task tracker primes the anomaly detector with what to expect. Affect shifts everyone's priorities before any of them start. The Cognitive Resolution Protocol is an application-layer protocol (OSI Layer 7, same as HTTP) that defines how an AI system's internal cognitive modules collectively interpret a single observation. Where MCP defines how agents access tools and A2A defines how agents talk to each other, CRP defines how an agent thinks about what it sees. It's the perception layer of the agentic AI stack. The key insight — one that forced a revision from v0.1.0 to v0.2.0 before publication — is that perception has two tiers, not one. Think about the human ear. It's not one thing. The cochlea converts sound waves to electrical signals. The auditory nerve decomposes frequencies. Brainstem nuclei handle sound localization. The auditory cortex does pattern recognition. And the vestibular system — physically part of the same organ — handles balance. One organ, multiple sub-components, serving completely different cognitive consumers. CRP mirrors this with a two-tier architecture: Sensory systems (Layer 0) receive broadly and produce rich, general-purpose output. A Visual Sense reports color distribution, layout structure, element density, and change detection. A Textual Sense extracts entities, detects patterns, estimates information density, and identifies language register. These senses are dumb but rich — they don't interpret, they report. Cognitive modules (Layer 2) consume sensory outputs for specialized interpretation. A Safety module reads the Textual Sense's entity extraction and decides whether PHI is present. A Task module reads the same output and matches workflow patterns. An Anomaly module reads it and checks for deviations from expectations. One sensory output, multiple cognitive interpretations. This separation means adding a new cognitive module for a different industry requires zero new sensory infrastructure. A Billing module for finance just consumes the same Textual and Contextual senses that a Clinical module uses in healthcare. The second revision that changed everything: memory is infrastructure, not a module. In v0.1.0, Memory was listed alongside Task and Anomaly as a peer that received broadcasts and responded to queries. But that creates a problem: cognitive modules need memory context to interpret observations, but memory needs the broadcast to know what context to provide. The fix is biological. Your hippocampus doesn't wait for requests. When you walk into a familiar room, relevant memories activate automatically — the context triggers retrieval. Memory isn't something you query. It's a lens that colors how everything else perceives. CRP v0.2.0 gives memory three roles in every cycle: it pre-loads relevant context before the broadcast (your cognitive modules receive history without asking for it), it responds to on-demand queries during the cycle (when the Anomaly module needs a specific historical comparison), and it forms new memories after the cycle (deciding what to store, how strongly, and how fast it should decay — based on salience and emotional weight). 1. The broadcast is a question, not an answer. The workspace doesn't classify the observation and send results to modules. It sends raw sensory context and asks "what do you see?" Each module interprets through its own lens. They can all be right simultaneously. 2. Your AI should feel before it thinks. An affective pre-processor generates emotional context — valence, arousal, urgency, novelty — before cognitive modules run. A screen full of red error indicators