Consequence closure

Abstract

Which physical systems are agents, and why? The sciences of life and mind presuppose an answerorganisms act, neurons learn, robots decide, software systems pursue goals yet no accepted theory states, non-circularly, the physical conditions under which a system is a locus of agency rather than a process, an aggregate, or an observer-imposed carving of a causal eld. Existing candidates either presuppose what they individuate (Markov- blanket and information-theoretic accounts, which require pre-given boundaries or parti- tions), answer a dierent question (organizational closure: persistence; causal emergence: scale selection given a partition), or are programmatic rather than formal (enactivist and organismic accounts). This paper proposes Consequence Closure Theory (CCT): an agent is a self-verifying and self-executing locus of consequence closurea macro-variable whose do-interventions on the world return, at the same scale and over the same horizon, as infor- mation about the consequences of those very interventions. Two interventional information quantities, the generative and predictive consequence loops, are dened in the tradition of Ay and Polani's information ow and Janzing et al.'s causal inuence; their normalised agreement denes a graded measure of consequence closure; agents are local maxima of this measure over scales, macro-maps, and horizons, and numerically identical agents are its equivalence classes. The theory is shown to inherit the partition content of the Markov- blanket formalism while avoiding its individuation circularity, and to explain, as corollaries, results currently scattered across causal emergence, informational autonomy, empowerment, and causal agent discovery. Two exact worked examples are computed: a redundant con- sensus channel, in which the agent exists at the macro- but not the micro-scale, and an XOR-coupled system, in which eective information rises under coarse-graining while no agent existsshowing that consequence closure and causal emergence come apart. Four testable predictions, seven steelmanned objections with replies, ve explicit falsication conditions, and the theory's stated boundaries (it is not a theory of consciousness, norma- tivity, or life) are provided. The paper is oered as a formal proposal for hostile scrutiny, not as an established result.