From Hallucination to Auditability: Solving the AI trust crisis through defensible architecture

Organisations are deploying AI into legal, financial and regulatory workflows faster than they are deploying the controls those workflows require. Many of the architectures being adopted — particularly those built exclusively on large language models — were never designed to produce the verifiable facts, transparent reasoning and reproducible outputs that consequential decisions demand. This paper argues that trust in AI is not a function of model sophistication, accuracy or fluency, but of defensibility: the ability to demonstrate after the fact what data a system used, what reasoning it performed, and why a particular output is reasonable. Defensibility is an architectural property, not a feature of the model. The paper sets out the architectural choices that distinguish defensible AI systems from indefensible ones — discriminative inference under version-locked weights and calibrated thresholds; Retrieval-Augmented Advisory workflows that cite authenticated source material verbatim; and Human-in-the-Loop governance that confines generative models to advisory rather than agentic roles — and proposes Bitemporal Chain of Custody™ as a reproducibility standard. The standard binds input data state (with valid and transaction time), model artefact, feature vector, context snapshot and decision trace into a single cryptographically verifiable audit record. Acceptance is defined by a reconstruction test: a third party holding only the audit record must be able to re-derive the recorded output and verify every hash, or the audit fails by construction. The paper maps these commitments to obligations under the EU AI Act (Articles 12–15), the UK regulatory framework (DSIT, ICO, FCA RTS 6, SM Pyrrho), and offers the standard openly for critique, refinement or supersession by formal standards bodies.