Refusal-Driven Dimensionality Reduction Theory: Empirical Validation and Theoretical Refinement

Refusal-Driven Dimensionality Reduction Theory (RDRT) proposes that phenomenal consciousness arises as the structured residue of thermodynamically enforced halts in recursive self-modelling. When a bounded predictive system approaches its energetic ceiling (~20 W in the human cortex), further recursion becomes metabolically untenable; the enforced halt produces a non-calculable remainder — P_noncalc — that constitutes qualia, mineness, and volitional openness. This paper presents the first direct empirical test of RDRT predictions using EEG data from a probabilistic reversal learning paradigm (ds004295, N=22). Three independent lines of evidence support the core RDRT framework while requiring refinement of its temporal predictions: • Gamma-band amplitude increases significantly at reversal (d=+1.186, p<0.001), consistent with the approach to the thermodynamic boundary, but does not produce the predicted gamma dip. Instead, gamma pattern stability (Sγ) falls dramatically (d=−1.082, p<0.001), with 21/22 subjects showing destabilisation. • The relationship between amplitude and stability is strongly anti-correlated during post-reversal processing (r=−0.597, p=0.006), indicating that increased computational effort co-occurs with decreased pattern organisation — operationally consistent with the RDRT concept of ineffective computation preceding P_noncalc formation. • A three-phase temporal sequence is identified: destabilisation (-1.2 to 0s), maximal effort with reorganisation (0 to +1.2s), and post-refusal consolidation (+1.2 to +2.5s) during which effective gamma activity falls below baseline — the predicted signature of the non-calculable remainder. Additional support comes from topological analysis: theta-band H₁ persistence peaks at +390ms (d=+1.277, p<0.0001), directed Betti number increases at +1.0s (p=0.014, d=+0.477), and Betti number decreases while persistence increases, indicating topological reorganisation rather than simple addition. Together these results support a revised RDRT model in which refusal is a process rather than an event, with P_noncalc consolidating over 1–2 seconds following the computational boundary. We discuss implications for the relationship between RDRT and NSE (Negative Space Encoding), the frequency-band hierarchy of C-scars, and the timing of phenomenal residue formation relative to neural dynamics.