Never Why: Causal Opacity and the Architecture of Honest AI Governance

Never why — not as a privacy constraint, not as a product decision, but as an architectural commitment to honest evidence. That is the thesis of this paper, and it belongs in the first paragraph so the reader understands what the argument is building toward: governance infrastructure that records only what can be cryptographically anchored and independently verified, and is honest about what cannot. Governance frameworks for agentic AI systems increasingly demand that those systems explain why they took a given action. This paper argues that demand is structurally impossible to satisfy — not because AI systems are dishonest, but because the causal process that produced any given output is not accessible to the system that produced it. When a large language model generates an explanation of its own reasoning, it is producing a plausible narrative using the same generative process that produced the original action. It is not retrieving a causal record. The explanation is a confabulation in the technical sense: generated after the fact, indistinguishable in form from the action it purports to describe, not grounded in an inspectable internal record of what actually happened. The governance consequence is worse than a gap. A governance system that accepts model-generated explanations as audit evidence is not building a compliance record. It is building a liability instrument — documentation that carries the appearance of accountability and collapses under cross-examination. That is the confabulation audit trap, and it is the direct operational consequence of misunderstanding what models can and cannot report about themselves. Mechanistic interpretability research — including Anthropic's published work on circuits, features, and superposition — has made meaningful progress toward understanding how specific capabilities emerge in transformer models. It has not produced, and does not claim to produce, the per-action causal record that governance frameworks require when they ask an AI system to explain why it acted. This paper engages that research directly and precisely, not to dismiss it but to establish exactly where it stops short of the governance claim. Honest governance infrastructure must limit its evidentiary claims to what can be cryptographically anchored and independently verified: who acted, when, where, and how — and assign interpretive authority over intent exclusively to accountable humans. This paper develops that argument, identifies the governance theater problem created by why-attribution systems, and describes the three-layer architecture — APR-Shaper, APR-Guard, APR-Lite, developed by Soft Armor Labs — that demonstrates the architecture is not theoretical. It has been built.