On responsibility, sovereignty, and the conditions under which delegated intelligence can be meaningfully governed.
v1.0 · 28 April 2026In any system of delegated intelligence, responsibility is not eliminated through delegation; it remains anchored in the human and must be explicitly structured across the system.
The human shifts from tactical execution — the "how" — to constitutional authorship: defining the objectives, constraints, and conditions under which the system operates. The agent functions as an extension of human intent within these defined structures, not as a replacement for human authority.
In architected coordination, the primary mechanism for defining control and legitimacy is the Specification — the governing structure of purpose, boundaries, measurement, and adaptation.
When the agent operates outside the specification, it has exited its authorised scope. Such deviations require containment, correction, or revocation of delegated authority.
Responsibility requires the ability to observe and evaluate. The specification must therefore ensure Epistemic Access: human-legible visibility into how the system operates relative to its governing constraints.
This includes, where feasible: traceable decision pathways, observable adherence to constraints, reproducibility under defined conditions, and clear evaluation signals linking outputs to defined criteria.
This verification bridge enables meaningful oversight. When it degrades — through opacity, untraceable behaviour, or loss of evaluability — the Architect's ability to exercise informed responsibility is diminished, and effective control degrades.
In systems of increasing complexity, governance cannot rely solely on continuous external oversight. It must be embedded in the structure of the system itself, making non-compliant behaviour detectable, constrained, and recoverable.
The Architect's primary responsibility is to design an enclosure: a governed environment shaped by:
This enclosure is not static; it is continuously evaluated and refined through structured audit and adaptation processes.
The responsibility that persists with the Architect extends to the quality of the specification itself. A PromptQ-assessed specification score is the primary instrument for detecting structural incompleteness at authorship time.
An Architect who deploys a system governed by a specification with identified structural gaps — particularly absent success definitions, unresolved internal conflicts, or missing quality gates — without addressing those gaps has not fulfilled the responsibility this philosophy requires.
Structural incompleteness in the specification is the most common and most tractable failure mode. The Architect's Vow includes the obligation to ensure the governing structure is structurally sound before delegating authority.
I am responsible for the system I define.
My agents operate within the constraints I establish.
I judge not each individual act in isolation, but the integrity of the governing structure.
I maintain epistemic access so that my authority remains meaningful and exercisable.
When visibility degrades or constraints fail, I act to restore control rather than assume compliance.
This philosophy, the governing principles, and the audit protocol together form a unified system: the philosophy defines responsibility, the principles define structural requirements, and the protocol governs continuous enforcement and refinement. None is sufficient in isolation; together they enable sustained, accountable control of agentic systems.
The framework adopts an epistemic stance of withholding final assent (epoché): evidence informs action and evaluation, but all conclusions remain provisional and subject to continuous audit, challenge, and refinement.
No specification or evidence system can fully capture system behaviour; skilled human evaluation remains necessary to interpret, challenge, and refine the governing structure. This philosophy does not guarantee correctness. It establishes the conditions under which responsibility can be meaningfully exercised in systems of delegated intelligence.
This framework addresses single-layer delegation. In multi-layer delegation chains, the evidence chain and responsibility attribution require additional structural provisions not defined in this version. Architects operating multi-layer agentic systems should treat each delegation layer as a separate governance scope requiring its own specification assessment.
Tibebu (arXiv:2604.07778, 2026) proves that for multi-agent collectives exceeding an accountability horizon, no framework can simultaneously satisfy attributability, foreseeability, non-vacuity, and completeness of responsibility assignment. This framework operates within the accountability horizon, not beyond it.