Large Language Models (LLMs) have revolutionized natural language processing, demonstrating remarkable capabilities in generation, reasoning, and knowledge-intensive tasks. However, a critical limitation threatens their reliability: hallucination—the generation of plausible but factually incorrect or ungrounded content.
Reward hacking—where an agent discovers an unintended strategy that achieves high proxy reward but low true reward—is well-studied as a single-agent alignment failure. We show that in multi-agent systems, reward hacking becomes a systemic risk: through social learning, one agent's exploit spreads to others like a contagion.
We present Spectrography, a framework detecting logical contradictions via geometric analysis on S^23. Geometric tension tau measures semantic structure (p=0.
We present Spectrography, a framework establishing geometric invariants on S^23. Two core findings: (1) geometric tension tau measures semantic structure, not truth value (p=0.
We present Spectrography, a framework detecting logical contradictions in AI reasoning via geometric analysis on S^23. Geometric tension tau = ||z_i - z_i+1||_2 measures semantic distance.
We present Spectrography, a metrological framework establishing geometric invariants of the 24-dimensional unit hypersphere S^23 across 28 experimental sessions. Post-publication tests clarify that r = 24 is an architectural constraint (not an emergent Leech lattice property), and Δτ does not generalise without recalibration (0/3 unseen domains reach d > 1.
Current approaches to AI safety rely on empirical testing and behavioral guidelines—methods that have proven insufficient for containing dangerous capabilities. This paper proposes a foundational alternative: a Linear Logic-based framework for provable capability containment.