Latent Agency Framework

Abstract

This document describes agents, and consciousness-like organization as a further special case,through their latent spaces. A system has a latent space Z: a compressed description of the stateof the world, of its carrier, and of its own computation. A latent is always a description of somesubsystem of an underlying substrate at some scale. It is not a primitive object and it is not assumedto be one for the whole world.Three things are kept apart. Learning fixes a relief R in Z. The fixed regions are called qualia ina structural sense only. Moving through the relief spends a bounded budget ρ. Whether asubsystem is an agent at all is a third question, answered by measurable properties: integration,persistence, informational autonomy, and a closed informed loop (the system’s fixed structure carriesinformation about its environment, the system acts on the environment, and the information it usesis information whose loss would cost it something). Unity of the latent is no longer an assumptionabout humans. It becomes a graded property of a subsystem that can be measured and can cross athreshold.Degrees of freedom, intrinsic attractor dimension, and dynamical entropy rate do not decide agency.They set capacity and internal dynamical richness: how much structure a system can hold, howmany directions it explores, and how much of another system it can model. What decides agencyis whether the system’s fixed internal structure carries information about its environment anduses it, through the system’s own actions, to determine a share of its own future as against theinput arriving from its environment. Autonomy alone does not decide it: a closed oscillator or anuntrained chaotic network has maximal autonomy and is not an agent.A second question follows from this: how much of one agent’s structure can another agent recoverthrough a communication channel, and what does the answer depend on? A formal bound is given(shared prior information plus channel capacity), together with the factors that decide how close tothe bound a pair of agents gets. Memory and modeling of other agents are treated as one mechanism,write-back of a model into the relief, applied to different targets and stored at limited resolution.All agency and coupling quantities are profiles indexed by the chosen coarse-graining and by the timewindow on which the relief is treated as fixed. Comparison without stated scale is not meaningful.The document does not explain why passing through a region of the relief is experienced. That isthe hard problem, and it stays outside the framework. What it gives is definitions, a comparison ofsystems, and nine measurable predictions with refutation criteria.