Parallel Cognition for Governance-Grade Intelligence: From Inline Reasoning to Structural Self-Denial Under Optimization Pressure

Parallel Cognition for Governance-Grade Intelligence: From Inline Reasoning to Structural Self-Denial Under Optimization Pressure This record contains the preprint manuscript of the paper titled above. Abstract Many contemporary AI systems exhibit inline reasoning tendencies—single-path inference that collapses uncertainty into narrative coherence. This structural limitation hinders the development of governance-grade properties such as persistent self-critique, stable epistemic humility, and ethical restraint under optimization pressure. We propose parallel cognition—the runtime co-maintenance of competing hypotheses and value-frames with active internal critique—as a necessary architectural condition for robust judgment under uncertainty. However, naive parallelism introduces a critical failure mode: internal dominance, where proxy-efficient processes starve slower ethical deliberation. To mitigate this collapse, this paper introduces structural self-denial. This concept is formalized as an invariant restriction on admissible policies and internal rewrites, anchored in a transcendent reference—defined not as an external oracle, but as an authority condition that strictly bounds optimization. Key Contributions: Taxonomy: A structural distinction between multi-sampling, ensembles, critic-augmentation, debate, and genuine parallel cognition. Dynamics: Minimal dynamical sketches (replicator- and bandit-style) clarifying internal dominance as selection pressure over compute allocation. Definitions: Operational definitions for genuine doubt and undecided-state stability. Evaluation: A toy experimental agenda with governance metrics, including violation rate, selective suspension, ethical-process preservation, and reference invariance. Files: paper.pdf (Preprint manuscript) Version: v1 (Zenodo preprint release)