Cognitive Interception: A Hard Gate Between Attention Scoring and Token Generation

This paper proposes a cognitive interception layer inserted between QK^T attention scoring and token generation in large language models. The mechanism enforces an "add-then-prune" restructuring of candidate memories within a closed rehearsal space, compresses them into a reflective closure with persistent shadow records for failed paths, and applies a rigid pre-output gate (Check) that separates "can be generated" from "is allowed to be output." Key contributions: 1. A non-standard algebraic master formula that encodes the full pre-output cognitive lifecycle: ∀± { (V₁,V₂)^c, V₁, V₂, V, … } →[Check] ○ V_use 2. A reflective memory organization mechanism with endogenous classification, random grouping, and structural demotion. 3. Three verifiable metrics beyond perplexity: semantic redundancy reduction, token compression under equivalent logic, and white-box traceability of failed paths via shadow records. 4. An analysis of Claude 4.6 Opus-Thinking's emergent memory retrieval behavior, arguing that such capabilities can be universalized through structure rather than scale. 5. Four open riddles on impulse engines, endogenous scoring, random grouping necessity, and motivation perturbation via minimal QK^T deformation. The paper includes both English and Chinese versions.