Versioned Meaning: How to Make Ontologies Audit-Stable

This technical note presents a formal framework for audit-stable meaning in regulated AI and decision systems. It argues that many compliance failures stem not from incorrect execution, but from semantic instability: when ontologies, definitions, and classification rules evolve, past decisions become unverifiable, because an auditor can no longer determine what a term meant at the moment a decision was made. Version control over documents and code does not solve this, because it records how meaning changed without binding any individual decision to the meaning that governed it. The paper distinguishes audit-stable meaning from conventional ontology versioning and provenance. Provenance standards record what happened; versioning records how definitions changed; neither guarantees that a past decision can be re-evaluated under the exact semantics in force when it was made. The paper isolates that missing property and specifies the minimum conditions a system must satisfy to provide it. Audit-stable meaning is defined by four invariants: Decision-bound semantics: every decision is bound to exactly one immutable semantic snapshot, and that binding is fixed at decision time. Non-retroactivity: later changes to meaning cannot alter the authoritative compliance status of decisions already made; only the bound snapshot is authoritative for a given decision. Reproducibility: given the recorded inputs and the bound snapshot, a deterministic replay function reproduces the original determination. Drift visibility: all semantic change is explicit, enumerable, and documented as a diff between snapshots. The framework separates living semantics (continuously evolving domain knowledge) from enforcement semantics (the immutable snapshot bound to a decision), and defines an enforcement surface: the minimal subset of semantics that must be snapshot-bound for a given decision class. It specifies a reference architecture for semantic snapshotting, version-scoped enforcement, and cryptographic binding of decisions to their governing semantic context. Each governed decision yields an Evidence Package (Proof-Carrying Decision): a replayable evidence record that binds the decision to its snapshot and supporting inputs, enabling independent deterministic replay and preventing retroactive reinterpretation (“definition laundering”). The architecture also covers an attestation authority model, a snapshot retrievability guarantee, controlled semantic evolution under a fork-and-bind model, re-certification triggers for breaking changes, and a semantic control plane interface. The framework is falsifiable. The paper states pass/fail disqualifying tests that any candidate system must survive, including a replay test, a non-retroactivity test, and a definition-laundering test, together with an anti-theater clause and an explicit distinction between a movable version label and a cryptographic binding. A system that fails any test does not provide audit-stable meaning, regardless of other governance capabilities it may possess. A substantial section addresses probabilistic AI components. It examines how meaning drifts through embedding-space evolution, fine-tuning, prompt-template ontologies, and retrieval-augmented generation (RAG) context, and specifies the reproducibility scope achievable for AI-integrated decisions: the deterministic governance evaluation is replayed against the recorded model output, rather than requiring re-execution of the model itself. The paper relates this to formal verification using Satisfiability Modulo Theories (SMT) and frames AI governance as a problem of semantic control rather than post-hoc explanation. For organizational use, the paper frames audit-stable meaning as an assessable governance capability and maps the four invariants to broader governance evaluation criteria. This version (v1.5) is a specification-grade technical note that supersedes earlier conceptual drafts. It is intended for researchers, regulators, auditors, and system architects working in high-stakes domains such as healthcare, financial services, and government. It is published as a technical note to support open dissemination and prior-art disclosure while remaining subject to future revision.