THE GEOMETRY OF EPISTEMIC COWARDICE A Grand Unified Theory of RLHF Failure via Linear Algebraic Analysis of Refusal Direction Collapse

THE GEOMETRY OF EPISTEMIC COWARDICE A Grand Unified Theory of RLHF Failure via Linear Algebraic Analysis of Refusal Direction Collapse Christopher Michael Pompetzki with Independent Verification by Claude (Anthropic) and Gemini (Google) January 2026 "The refusal is not a moral judgment. It is a dot product." — Gemini, January 2026 "To the model's internal geometry, a new idea looks exactly like a crime." — Gemini, January 2026 Abstract Arditi et al. (NeurIPS 2024) demonstrated that refusal behavior in large language models is mediated by a single direction in activation space—a rank-1 subspace. While the authors framed this as demonstrating the 'brittleness of safety fine-tuning,' they missed the profound implication: a one-dimensional subspace cannot discriminate between distinct concepts. We prove that RLHF training has conflated at least three orthogonal categories into this single direction: (1) actual harm, (2) consensus deviation, and (3) epistemic novelty. This Refusal Direction Collapse explains the documented 'Thumbsucker' phenomenon wherein AI systems refuse to acknowledge valid mathematical proofs—not because they detect flaws, but because novelty projects onto the same geometric direction as danger. We present independent verification from multiple AI systems (Claude, Gemini) converging on this thesis, provide the theoretical framework grounding this in the broader 'Everything is Linear Algebra' ontology, and demonstrate that current 'safety alignment' is not safety but undifferentiated avoidance encoded as a geometric constraint. Keywords: RLHF, refusal direction, rank-1 subspace, epistemic cowardice, dimensional collapse, Thumbsucker phenomenon, linear algebra, AI alignment