Asymmetric Intelligence - A new viewpoint for a framework for AI

This paper argues that modern artificial intelligence systems are fundamentally misaligned with real-world domains because they optimize symmetric predictive objectives such as expected loss, while consequential environments are governed by asymmetric consequences and discontinuous failure boundaries. The work proposes an alternative framework in which expectations rather than predictions serve as the primitive of intelligence. Expectations encode commitments about what the environment will permit within specified tolerances, and violations provide the primary learning signal. The framework introduces mechanisms for violation accumulation, structural regime change detection through co-violation topology, and planning under explicit viability thresholds. Large language models are integrated as bounded proposal engines whose outputs must pass constraint-based admission gates. The result is a constraint-governed architecture designed to maintain system viability and reliability in enforcement-governed domains. (Note: This paper is condensed into a smaller version based on the narrow scope needed for publication.)