When a developer changes a security-sensitive component, an AI assistant increasingly proposes the security controls the change should satisfy. In regulated or safety-critical work, that proposal must come with an explanation that can be audited, reproduced, and checked against its sources. The usual way to produce one is post-hoc explainable AI: a separate step reconstructs, after the fact, why an opaque model produced its output. That step inherits a faithfulness gap — the reconstruction need not be what actually drove the output. We argue that, for security guidance, the gap comes in part from asking the wrong question. The thing to be explained — the explanandum — is not why the model emitted a token, but why a given obligation must hold for this code, this change, this regulation. That is a normative "why", not a causal one. When the assistant is grounded in a bounded, typed, source-traceable application-security ontology, the explanation is not reconstructed after the decision. It is ante-hoc: its content is the same typed obligation graph that constrained the recommendation in the first place. Grounding does not close that causal gap; it removes the gap between the explanation and the recommendation it explains, by changing what the explanation is about — and relocates the remaining burden onto the grounding's own coverage and provenance, which we make explicit. We give a semantic account of that explanation. We say what it explains (a deontic obligation), how it is structured (as a Toulmin argument — claim, grounds, warrant, backing — with provenance as the warrant), why it is faithful (by construction, not by approximation), and for whom (the same structure serves an engineer, an LLM agent, and an auditor). We state its limits plainly. The ontology, its substrate, its delivery mechanism, and its evaluation are companion work; our contribution is the account of the explanations they make possible.
Pedro Farinha (Thu,) studied this question.