Routing Uncertainty in AI Systems: From Failure Mode to Governance Signal

This paper proposes a paradigm shift in how autonomous AI systems handle uncertainty. Rather than treating high-entropy states as failure modes to be suppressed through calibration or ensembling, we formalize uncertainty as a high-fidelity governance signal that triggers tiered routing to a rights-based oversight layer. The framework introduces three core contributions: An Uncertainty Routing function grounded in Conformal Prediction, which uses prediction set cardinality to direct inputs along autonomous, precedent-retrieval, or governance-escalation pathways — with distribution-free coverage guarantees. An Irreversibility Coefficient that weights routing decisions by consequential severity across four dimensions (temporal irreversibility, scope of impact, rights proximity, and asymmetry of harm), ensuring that high-stakes decisions receive governance review regardless of model confidence. A Precedent-Based Value System with versioning, conflict resolution, and temporal decay protocols, enabling the system to accumulate resolved cases as vector embeddings — analogous to common-law precedent — while remaining responsive to evolving norms and axiom revisions This work is situated within the ELEANOR (Ethical Leadership Engine for Autonomous Navigation of Rights-Based Reasoning) V8 framework and the Intelligence Pluralism doctrine, which holds that no single intelligence — human or artificial — should be trusted with unilateral authority over consequential decisions. ELEANOR is complementary to training-time alignment approaches such as Anthropic's Constitutional AI, providing runtime governance for deployed systems making real-world consequential decisions.