Where the Theories Stop: Practical Limits of the Free Energy Principle and Integrated Information Theory in a Running Cognitive Architecture

This paper examines where two dominant theoretical frameworks in computational 
neuroscience — Friston's Free Energy Principle and Tononi's Integrated Information 
Theory — provide genuine engineering traction and where they stop, drawing on 
observational evidence from Anima, a running neuroscience-inspired cognitive 
architecture implemented in Julia.

We document specific cases where FEP-derived mechanisms (prediction-error-driven 
state change, recursive phi feedback, active-inference initiative) produce measurable 
behavioral effects, and cases where the phi metric as implemented is more decorative 
than functional. We introduce the concept of endorsement — whether expressed language 
is consistent with the system's internal state and causal ownership — as a behavioral 
signal that is more informative than phi alone, and not captured by either framework.

No claim is made that either theory "does not work." The claim is that both work 
differently than their strongest proponents suggest, and that honest engineering 
requires knowing the difference.

Presented as a research artifact with observational evidence from session logs. 
Not a peer-reviewed experimental result.

Project repository: https://github.com/stell2026/Anima