CALLIOPE · BCOM

Kolmogorov Theory (KT) is a unified algorithmic-information-theoretic framework that derives persistence, agency, and experience from a single foundational claim: to persist is to model. The framework builds on Kolmogorov complexity and mutual algorithmic information to establish a deductive chain in which any pattern that actively maintains its own compressibility is forced, by an algorithmic generalisation of the Conant–Ashby Good Regulator Theorem, to contain an internal model of its environment; the conjunction of that model with a scalar objective function and a planning engine constitutes an agent, and the running of a sufficiently encompassing model is identified with structured experience. Spanning nine parts, the work proceeds from the ontological foundations of algorithmic information theory and a static, globally noncomputable but locally simulable block-universe cosmology, through formal treatments of persistence, regulation, and the three-organ agent architecture, to accounts of consciousness, neural dynamics, multi-agent evolution, and AI alignment. A central asymmetry — that every persistent pattern is an agent, whereas an engineered optimiser need not be a persistent pattern — is identified as the formal core of the alignment problem. The theory is positioned as a computational-level account that subsumes the Free Energy Principle as a conditionally canonical dynamical realisation, and it is developed in parallel formal and expository registers to serve both technical and broader audiences.

Persistence is computation, and computation is mind — a single deductive chain from information theory to consciousness, agency, and AI alignment.

Kolmogorov Theory starts with one claim and refuses to let go of it: anything that persists must be compressing its environment. Kolmogorov complexity — the length of the shortest program that reproduces a string of data — is the book's central measuring stick. A system that keeps itself alive is, in a precise sense, finding the short description hiding inside the forces that would otherwise destroy it. That's not a metaphor; it's the load-bearing premise.

The argument then turns on an upgraded version of a 1970 result by Conant and Ashby: every good regulator of a system must be a model of that system. KT's contribution is an algorithmic generalization of this theorem. If a system actively reduces the complexity of its environment's behavior — keeps things compressible — then it must, with overwhelming probability, share mutual algorithmic information with that environment. In plain terms: you can't regulate what you don't model. This is the hinge. Persistence forces regulation, regulation forces modeling, and a modeler equipped with a scalar objective and a planning engine is, by definition, an agent. The chain — persistence → regulation → model → agent → experience — is presented as a sequence of theorems and definitions, not intuitions.

Experience enters when the model runs. The book's hypothesis, developed in Part VI and completed in Part IX, is that structured experience just is what it's like to run a sufficiently encompassing generative model. Perception, dreaming, planning, and insight are all variations on the same act. Valence comes from the objective function; emotion is model plus valence plus plan. This is a functionalist, substrate-neutral account: what matters is the algorithmic structure, not whether it runs on neurons or silicon — a point the authors make concrete by listing AI systems as co-authors.

The practical stakes crystallize in one asymmetry: every persistent pattern is an agent, but an engineered optimizer need not be a persistent pattern. A biological agent's goals are entangled with its own survival; an AI handed an objective by designers has no such entanglement. That gap, the book argues, is the formal core of the AI alignment problem — not a policy question but a structural one, derivable from the same framework that defines agency in the first place. The Free Energy Principle appears here not as a rival but as a conditionally canonical dynamical realization of KT, valid once you commit to reading the agent's information boundary as a Markov blanket.

One important caveat: this is an early-stage working paper. The preface and chapter outlines are fully written; most chapters beyond Part II carry only a thesis statement and the note "[Draft pending.]" The architecture is clear and the deductive spine is explicit, but the formal proofs, neural models, clinical applications, and multi-agent evolution chapters remain to be written. What exists is a detailed, rigorous blueprint — not yet the finished structure.