Structural Alignment in Practice

Regulatory frameworks tell organizations what they must document, oversee and report. Structural alignment asks a further question: does the continuity of reasoning, agency and accountability that those obligations depend on actually survive inside the system?

This page is the entry point to Audited State's structural readings of AI governance regimes. It sets out the six pillars that every reading uses, and lists the regimes read so far. For the concept itself, see Structural Alignment. For the argument in essay form, see Structural Alignment: Why AI governance is becoming an infrastructure problem.

A note on scope. These readings are structural translations, not legal advice. They do not interpret statutory text in a binding manner, offer a compliance checklist, or track amendments. Each regime is read at the level of its durable intent.

Why a structural reading?

Governance has traditionally assumed that reasoning can be reconstructed after the fact. In distributed AI systems, where reasoning passes across agents, workflows and memory layers, that assumption often fails. Traditional alignment asks whether a model produced acceptable behavior. Structural alignment asks whether accountable reasoning survives every transition the system goes through.

The term is used here for a property of systems, not of legal texts. It is unrelated to harmonizing rules across jurisdictions, and to other uses of the same words elsewhere, such as comparing protein structures or analogies.

The six pillars

Each regime is read through the same six pillars, so that readings can be compared. Each pillar is stated below together with the question it puts to any regime.

Structural Embedding

Alignment as a property of how the system operates, not only of what it outputs. Continuity is built into the operational structure, so it does not depend on inspecting outputs afterwards.

The question: Is continuity part of how the system works, or something reconstructed from its outputs when someone asks?

Decision Lineage

The traceable continuity of intent, from the moment a possibility is taken seriously to the moment a decision becomes visible. It is a chain of reasoning, not a log of outcomes.

The question: Does the documentation preserve the reasoning behind a decision, or only the fact that it was made?

Continuity of Agency

A person's unbroken understanding of the system's actual state over time: oversight that does not lapse between one decision and the next.

The question: Does human oversight stay current between checkpoints, or only at them?

Audit and Repair

Recognizing the divergence between intention and outcome, and continuing with revised understanding while the historical record stays intact.

The question: When something goes wrong, is the prior state preserved and learning carried forward, or is the record overwritten?

Stewardship of Structure

Custodianship of the reasoning infrastructure itself: responsibility not for any single decision, but for keeping the mechanisms trustworthy, held at arm's length from the decisions they check.

The question: Who is answerable for whether the mechanisms themselves remain sound?

Visibility

Eliminating untraceable determination in autonomous execution, so that what a system decides on its own remains traceable to where and how it was decided.

The question: Can every determination made in autonomous execution be traced back to its origin?

Regimes read so far

EU AI Act

The AI Act's obligation clusters, read as structural requirements:

US federal AI policy

A discussion draft, a national-security memorandum and an executive order, read through the six pillars, with the overlaps and the gaps stated.

Why this matters

Compliance can satisfy a requirement while the reasoning behind a system has already collapsed. The pillars give a way to tell the difference, whatever regime is being read. Audited State proposes that compliance becomes durable only when continuity is architectural.

Related