Root Knowledge
Embodied Knowledge as a Foundation for Coherent Human–AI Interaction
Neuroscience increasingly explains the body, brain, and self through interoception, prediction, embodiment, and spatiotemporal accounts of psychopathology — yet it remains unclear how these layers interact to produce a coherent self-narrative. In parallel, artificial intelligence has advanced to large-scale generative systems capable of organizing information and producing precise outputs, but remains largely disconnected from how such outputs are integrated by the human system. This paper proposes a coherence-centered model of human-AI interaction in which AI is designed not only to generate faster or smarter output, but to scaffold information in ways that support meaning-making, self-narrative continuity, and integration across physical, biological, neural, and symbolic layers.
Intelligence begins in the body
Before modern computation, intelligence was already functioning as prediction through structure. Friston formalized this as a unifying principle — that perception, action, learning, and biological self-organization may all be understood as continuous prediction error minimization across hierarchical levels of the nervous system. Yet a central gap remains: how information becomes meaningful, integrated across time, or experienced as a coherent lived reality is not fully explained by output-level models.
The body is not a passive receiver. Interoceptive signals — heartbeat, breath, visceral tension — actively shape emotion, decision-making, and self-awareness. The self is not a fixed internal object but a temporally organized process shaped by prediction, bodily regulation, and ongoing interaction with the environment.
Figure 1. The Root of Knowledge: A Limbic-Interoceptive-Foundational Framework. A layered view of embodied knowing, moving from bodily grounding and interoceptive signaling to emotional filtering and meaning-making. Adapted from Avanzo (2026a).
Fragmentation as cross-layer misalignment
Modern conditions amplify fragmentation. Information overload, digital saturation, sleep disruption, and chronic uncertainty place sustained pressure on regulatory systems. Research increasingly suggests that many mental health conditions involve disturbances in large-scale brain dynamics, interoceptive processing, temporal continuity, and self-modeling rather than isolated cognitive deficits.
Fragmentation reflects dysregulation across multiple interacting layers of the human system — not a failure of thought, but a failure of coordination across body, neural dynamics, and self-narrative. The felt sense of continuity depends on successful cross-level coordination.
Figure 2. Self-amplifying cycle of integrative instability. Sustained allostatic load propagates downward across regulatory levels, reducing autonomic flexibility, increasing prediction error, and degrading the coherent self-model. Adapted from Avanzo (2026a).
Current AI amplifies what it should reduce
Current human-AI interaction is largely organized around efficiency metrics: faster answers, task completion, personalization, and output optimization. These advances are valuable, yet they primarily treat intelligence as production rather than integration. AI systems are not typically designed to support how information is interpreted, embodied, and integrated into lived experience.
Current model. Epistemic misalignment. No feedback loop between output and human state.
Root Knowledge model. Coherence-centered architecture. Integration layer prioritizes epistemic alignment over engagement.
AI as epistemic scaffold, not output generator
This paper introduces an intermediate integration layer that functions as an informational scaffold between the human user and large-scale data systems. Rather than optimizing primarily for engagement, this layer prioritizes epistemic alignment: the fit between incoming information, contextual relevance, and the user's current goals, state, and meaning structures.
In predictive processing terms, such a system attempts to reduce unnecessary uncertainty at the level of interaction by organizing information in ways that are more interpretable, relevant, and temporally sequenced for the human user. When the integration layer is designed for coherence rather than engagement, prediction error decreases — not only in machine outputs, but in the cognitive load placed on the human receiving them.
What this work does not claim
This framework remains in its early conceptual stages. The theoretical architecture and design principles proposed here have not yet been empirically validated, and claims about effectiveness await systematic testing. What is offered here is not a finished system, but a structured proposal: grounded in established research, internally consistent, and designed to generate testable predictions.
A companion paper introduces the M-RFT metric — a preliminary mathematical index designed to estimate coherence across layers of the human system. The N=1 longitudinal study using the Root Extension application provides preliminary feasibility data. Both are embargoed until June 1, 2026 and available upon request for academic or research purposes. Full empirical validation requires multimodal data collection combining physiological, neuroimaging, and narrative measures — the target of the planned N=20–40 validation trial.