Two Kinds of Complexity and Distributed Brakes Heterogeneous Homeostatic Loops, Same-Source Correlation and an AGI Outside Evolution
Year: 2026
Type: preprint
Abstract: It is commonly assumed that a sufficiently intelligent agent comes with its own constraints. Evolution answers otherwise: constraints are not built into intelligence but are a second system that co-evolved with it. This paper puts that intuition into a checkable form. First, there are two kinds of complexity, neither implying the other: parametric or fitting complexity, and heterogeneous homeostatic differentiation—how many loops inside a system each have their own set point, their own reading channel, and can veto the main action on their own. Evolution grew the two roughly together; large models and AGI are the first large systems to grow along the first axis alone. Second, “the more complex, the more centralised” is imprecise: by Chapter Fourteen of the book and How Many Layers, centralisation is at least four separately adjustable quantities, the number of layers is set by the spectrum of screening lengths of environmental disturbances, conjunctive functions must be concentrated, and within-layer concentration has an interior optimum; rising complexity moves the location of the optimal concentration, while the brakes are distributed across the layers. Third (tier B), brakes aggregate by “any one can stop”, so the bound 1/ρ does not apply; the conditional miss when trouble comes depends on two correlations—between brakes, and between brakes and violations—and on shared blind spots. With demonstration parameters (violation rate 1%, ordinary miss of a single brake 10%), at same-source correlation 0.6 a single brake misses 0.65 of real violations and eight same-source brakes still miss 0.18; if the added heterologous brake is wholly independent of the violation, one same-source plus one heterologous brake falls to 0.065. But this advantage is extremely sensitive to the heterologous brake’s own correlation: a loading of only 0.10 on the violation factor brings “one same-source plus one heterologous” level with eight same-source brakes, because violations are tail events and tails amplify small correlations. The design implication must therefore be written as a conditional on both correlations, and heterology must be measured. Fourth (tier B), biological brakes are self-paid, own their readings and lie beyond the reach of the main optimiser, whereas external brakes are maintained and rewritable, are eroded one by one in order of “the delay of keeping them ÷ the cost of removing them”, and may be switched off together by one switch; intelligence itself turns self-paid brakes into maintained ones. Fifth (tier B), disagreement is the bill for brakes: the couplings that produce conflict are the couplings that produce vetoes. When each brake is calibrated separately, lowering their correlation lowers joint blindness and raises joint false alarms; when joint false alarms are calibrated to a common level, lower correlation is better on both counts, and the bill reduces to the cost of adjudicating disagreements—while joint calibration is itself a shared switch. Sixth (tier B), the cleansing paradox: using one AI to remove reading conflicts among many people raises the correlation between their readings; in the demonstration, the effective number of independent readings among a thousand people falls from about 9.9 to about 1.8. Seventh, a mechanism proxy of R-1 and R-2 has received a preliminary execution: on existing annotations of real LLM guardrails, the lift with which two models from the same developer fail together is 20.4, against a median of 4.9 across developers, in the same direction as existing large-sample studies; on controlled monitors facing disguised attacks against the primary guard, the conditional miss of same-substrate second-line monitors when the primary misses is 0.94, that of fully heterologous ones 0.34, although the latter are worse on their own—second-line brakes should be chosen by their conditional miss when the primary misses, not by stand-alone accuracy; facing topics no brake has seen, changing features or data does not help, and joint misses stay above 0.8. The disguised attacks use only four templates; robustness checks by template clusters, rotation of the primary guard and multiple seeds have their criteria fixed and remain to be run, and one deviation from the pre-registration is corrected here. From these the paper derives conditional design criteria for an external brake layer, connects them item by item to external work on homeostatic physiology, requisite variety, biological degeneracy, common-mode failure and difficulty correlation in multi-version software, conditional dependence of diagnostic tests, correlated errors of large language models, veto-player theory, the multiple hits of cancer and algorithmic monoculture, and registers five predictions testable on today’s AI systems and institutions; R-1 and R-2 have received a preliminary execution on public data in the form of a mechanism proxy, the other three have not yet been executed.
Author: Li, Qinfu
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225903076
