Learning Meaning: An Ontology-Learning Mechanism for Governable Specific-Purpose AI

Governable specific-purpose AI systems produce typed, evidence-linked outputs under constitutive audit. This working paper introduces an ontology-learning mechanism that turns such governed operation into the sensor of semantic learning: every governed analysis emits typed claims; claims become ontological candidates; a human curation gate consecrates terms; each versioned ontology release recalibrates the next analysis. The contribution is threefold. First, a method: a constitutional life-cycle for domain terms (candidate, observed, curated, consecrated, deprecated) with hard rules — extraction over decree, mandatory human curation, golden-set gating, and hash-versioned releases declared in every trace. Second, an architecture: a five-stage pipeline (harvest, distillation, curation, release, feedback) that lives outside the platform core, feeds through a typed port, and is itself traced by the governance discipline it serves. Third, a formal dynamics: distinguishing the ontology state Omega(t) from the learning flux dOmega/dt, the system is given a Lagrangian L = (1/2)·gamma·(dOmega/dt)² − V(Omega,t), and Noether's theorem shows that the architectural invariants of governable AI — provider neutrality, domain invariance, constitutive audit, and hash-versioned provenance — act as symmetries whose conserved quantities are precisely the governance properties the system must guarantee. The result reframes governance from a cost imposed on learning to the very instrument that makes disciplined semantic learning possible. We term the resulting programme governed ontology learning. This is a concept and methods paper (no experiments); it is the seventh entry in the author's White Paper Series on AI Governance and Responsible Acceleration. Files under embargo until 1 November 2026; CC BY 4.0 upon release.