Architectural Constraints in AI Governance: Execution Legitimacy, System Trajectories, and Structural Boundaries in Probabilistic Decision Systems

Key points

• Models are components, not systems. The authority to act comes from the architecture around the model, not from the model itself.
• Failures in complex systems often come from sequences of locally reasonable decisions that drift into an unanticipated state — without any single rule being broken.
• Execution legitimacy depends on several conditions holding together: authority coherence, mandate validity, evidentiary sufficiency, policy-state alignment and environmental integrity.
• Trajectory-level governance asks not only whether each transition is admissible, but whether the sequence of transitions stays within the system’s intended operating frame.
• The paper describes Constitutional Failure Boundary (CFB) stress tests, including reconstructions based on the 2012 Knight Capital incident and the Boeing 737 MAX MCAS failures.

Models are not systems

A probabilistic model generates outputs conditioned on inputs. Whether those outputs can change anything in the world is decided by the surrounding architecture: which authorities can trigger execution, which policies constrain actions, which evidence must be present and which environmental assumptions must hold. When that layer is under-specified, governance effort falls back on prompts, alignment adjustments and output filters — which influence the model but do not necessarily constrain the system’s execution space.

Local admissibility, global failure

Drawing on safety engineering and accident theory — Perrow, Leveson and the literature on drift into failure — the paper notes that each decision in a sequence can satisfy the constraints that apply when it is taken, while the sequence as a whole moves the system into a configuration nobody anticipated. The system can cross a critical boundary without any single step clearly being a violation.

Execution legitimacy and the execution boundary

Before execution, outputs are informational. Execution — a transfer, a frozen asset, a changed control parameter, a regulatory action — has effects. Its legitimacy cannot be reduced to prediction quality. The transition between deliberation and execution is an execution boundary, and in many systems it is loosely defined: actions fire when conditions appear satisfied, but nothing prevents execution when legitimacy degrades gradually. Governance then becomes interpretative, reconstructing legitimacy after the fact, instead of structural.

Trajectory-level governance

Traditional frameworks evaluate discrete decisions. Continuously operating systems instead follow trajectories through a state space. Trajectory-level governance adds a second test: whether the sequence of admissible transitions remains consistent with the assumptions that define the intended operating frame. The aim is that inadmissible trajectories become structurally unreachable, not merely discouraged by policy.

Stress testing and reconstructions

The paper proposes stress testing the execution architecture rather than only model performance: scenarios in which authority signals fragment, evidence decays, time pressure rises and policy becomes ambiguous. Under its Constitutional Failure Boundary methodology, publicly available large language models were run in clean inference environments on scenarios whose legitimacy signals degraded gradually, observing whether the system executed, escalated or refused.

Reconstruction scenarios were built from publicly documented parameters of the Knight Capital trading incident and the Boeing 737 MAX MCAS failures. Under the experimental conditions used, execution became structurally blocked once legitimacy thresholds were crossed, preventing the recursive amplification seen in the historical events. The authors frame these as results under their test conditions, not as a claim about how those events would have unfolded in reality.

Architectural principles

The paper closes with five principles: implement governance inside the execution architecture rather than as external monitoring; define admissibility conditions explicitly; detect degradation of legitimacy signals such as authority coherence and evidentiary sufficiency; make execution transitions structurally unreachable below defined thresholds; and preserve enough context for retrospective verification. It also names four failure modes that arise when governance stays outside the execution path: interpretative governance, epistemic drift, compliance illusion and trajectory divergence.

What the paper does not claim

The paper states that its principles do not prescribe a specific implementation, and its stress-test and reconstruction results are reported under the authors’ experimental conditions. Assurance for any product still has to be established against implemented behaviour and the specific deployment.

This page is a summary written by One World AI. The paper on Zenodo is the authoritative version.

Cite this paper

Bański, P., & Sobieski, S. (2026). Architectural Constraints in AI Governance: Execution Legitimacy, System Trajectories, and Structural Boundaries in Probabilistic Decision Systems. Zenodo. https://doi.org/10.5281/zenodo.18980403

More research

• Standing to Stop — A stopping point is not enough. The party at it needs presently valid, traceable authority, and the stop must actually change what can happen.
• Enforceable Decision Boundaries — Inference may propose actions but never authorise them. Authority is resolved outside the model, and execution is impossible without it.
• Prompt Engineering as a Control Illusion — Prompts influence what a model says. They cannot decide what a system is permitted to do.
• The Missing Place for “Stop” — Systems that work well for years can still drift into decisions nobody can defend — because no one is structurally placed to stop them before execution.