The Four-Model Theory of Consciousness: A Simulation-Based Framework Unifying the Hard Problem, Binding, and Altered States

Abstract

Consciousness, on the Four-Model Theory, is constituted by ongoing self-simulation: a system builds a model of itself, relates to that model, and interacts with it. Models divide along two axes — scope (world vs. self) and mode (implicit vs. explicit) — yielding four model kinds rather than four modules. The implicit models are substrate-level, learned, and non-conscious; the explicit models are virtual, transient, and phenomenal. Qualia are constitutive properties of the computational level, incoherent at the substrate level as a spreadsheet cell's value is incoherent at the transistor level, which reframes the Hard Problem as resting on a category error: phenomenal properties are sought at a level where they do not exist. The framework rests on three principles. Open-ended computation allows free modeling, and near-criticality is the signature that property leaves in neural tissue rather than the requirement itself. Recurrent non-linear dynamics of sufficient complexity host a second computational level: persistent activity patterns individuated by their mutual effects rather than by the cells that carry them. Phenomenology is a real, physical effect on that level, arising where the modeling closes on itself — where the self-model and the world-model each enter the other's update rule. That this closure constitutes phenomenality is argued for and not derived (Section 3.4.3). The four kinds are then a minimum the architecture forces rather than a posit: four is the floor, not the ceiling. Redirection of the modeling resource, variable permeability and forking of the self-model follow as consequences, and through them the theory accounts for psychedelic phenomenology, anesthetic mechanisms, dream and lucid-dream states, split-brain phenomena, dissociative identity disorder and animal consciousness. Several claims derived from the theory's 2015 axioms have since received converging independent support (Xu et al., 2024; Hengen & Shew, 2025; Algom & Shriki, 2026), though that convergence concerns the computational requirement and is shared with other criticality-based frameworks. Novel predictions — among them that psychedelics should alleviate anosognosia — remain largely untested. Changelog v16 **v16 (2026-10-02)** Supersedes v15 (2026-08-26). Theory additions, a literature update covering work published since v15, and a revision of how the criticality requirement and the in-silico evidence are stated. *Theory.* A third principle, mutual closure, is stated in the abstract, §3 and §3.4. §3.4.4 treats emotions as qualia and states the memory and time conditions for human-like consciousness. §4.2.2 adds a paragraph on free will. The definitions section now defines information as data to which a coding can be assigned, adds one clause at the definition of knowledge (knowledge can be false; holding it is a separate relation), and adds glossary entries for *reference system* and *knowledge*. §3.5 adds a sentence on the term *sentience*. *Literature.* Work published since v15 is answered or cited, among it the hot-zone replication and null results, the propofol results from deep sedation (reported as a null), Sipling, Zhang and Di Ventra on criticality, Tiehen and Tubert on superintelligent zombies, and global-workspace work on large language models. §8.7 now answers "In part" to whether the theory's predictions are unique. Each new citation was verified against its source; unverifiable candidates were dropped. *Criticality and in-silico evidence.* The requirement is stated as Class 4 computational capability; branching ratio and Lyapunov exponent are instruments for it, and a branching ratio is reported only beside a rate-matched surrogate. §8.9 reports the edge-peak in computational capability as a single run, separately from the banked results. *Corrections.* Five internal inconsistencies are resolved (§4.2.4 against §4.4, the Prediction 3 summary against §8.4, the status of Xu et al. 2024, one leg of the §8.9 trichotomy, a stale description of the intelligence model). Four redundant self-citations are removed. Two bibliographic records are updated to their published versions. The citations of the companion papers carry their current titles. The acknowledgements are revised. The paper is 160 pages.