Our Retrenchment Theory, outlined in my book Retrenchment, holds that observation carries inherent computational limits, and that the core nature of cognitive systems is continuous retrenchment. When the redundancy penalty γ exceeds its critical value (γ > ½), retrenchment manifests as a phase transition: an information group under finite resources bifurcates into complementary poles. The seed of bifurcation is the coupling of input differences with accumulated-cognition differences—regions receiving different inputs, filtered through different knowledge bases, diverge irreversibly. Perfectly identical copies carry no seed and do not polarize. Within a single brain, this is how brain regions form. Once regions coexist in the same resource-bounded system, they communicate; when coupling K sits at the synchrony edge, information gain is maximal. Polarization supplies complementarity, communication supplies integration, and their synthesis is tension T. Conscious states arise when γ > ½, K sits in the sweet zone, and T remains integrable. Each pole polarizes again by the same equation, recursively, in a self-similar fractal.
Grounded in this theory, we built DarwinSys with a heterogeneous architecture: two independent retrenchment units generate divergent world representations, predicting and cross-validating each other. Mismatched representations and prediction errors act as the native driving force of cognition. Even without external stimuli, internal perturbations persist; driven by this coupling and tension, an open-ended stream of conscious-like cognitive states arises. When external signals break old stable states, new cognitive possibilities emerge endogenously, driving cognitive leaps.
DarwinSys continues this universal logic artificially: two heterogeneous simplifiers predicting and cross-checking each other, where mismatch itself becomes the engine of cognition.