Cognitive Exoskeletons: External Structure as the Determinant of LLM Output Identity

抄録

Large language models produce impressive but inconsistent outputs. The field has tried two fixes: change the model (fine-tuning, RLHF) or change the human (prompt engineering). Both miss what we believe is the actual lever. We propose attaching a cognitive exoskeleton—an external structural layer that wraps around the AI and shapes its output character through architecture, not parameter modification. We define the term precisely: a cognitive exoskeleton must be model-agnostic, identity-preserving across model switches, applicable across modalities, and self-correcting. Two implementations—SPARQ for output quality and GHOSTY COLLIDER for creative leaps—have been deployed across ~5,200 retained outputs in text, image, and video. Our central observation: wrap different LLMs in the same exoskeleton and the outputs converge; wrap the same LLM in different exoskeletons and the outputs diverge. This has not been validated through blinded evaluation. Drawing on Maisano et al. (2026), who showed that attentional filtering structure determines creative problem-solving style, we propose a functional analogy between exoskeleton interventions and cognitive filtering. We advance an operationally observed claim—that external structure significantly shapes output consistency—and a stronger hypothesis requiring validation: that output identity is an emergent property of the exoskeleton, not the model.