Article II — Governance Before Execution

Execution is subordinate to governance. No action may precede classification.

---

Doctrine

Governance is not a layer applied after execution. It is the precondition for execution. Every action proposed by an AI system must be classified, evaluated, and authorized before it reaches infrastructure. The governance layer stands between proposal and execution — not as a filter on the output, but as the structural authority that determines whether execution proceeds at all.

Execution without governance is equivalent to actuation without control feedback.¹ A system that acts before governance evaluates it is not a fast governance system. It is an ungoverned system.

No action may precede classification. This is not a performance constraint. It is a constitutional requirement.

---

Meaning

The analogy to control feedback is precise. In control systems engineering, actuation without feedback produces instability — the system cannot correct for error because it has no mechanism to observe its own state relative to the desired outcome. An AI system that executes before governance evaluates it is in the same structural position: it cannot be corrected, cannot be attributed, and cannot be held accountable for what it has already done.

Governance Before Execution establishes that the governance layer is not advisory. It is not a logging mechanism. It is not a post-execution audit. It is the evaluating authority that stands between every action proposal and the infrastructure that would execute it. The governance decision — ALLOW, DENY, ESCALATE, REQUIRE_CONFIRMATION — is issued before execution. The tool proxy layer executes only what governance has approved.

Governance state defines the execution context: the active doctrine version, the operational mode, the threat posture, the authority context, the protocol enforcement rules, and the audit requirements. An action evaluated against the wrong governance state is an action evaluated incorrectly, even if it is evaluated at all.

---

In Practice

The AEGIS system stack enforces this doctrine structurally. The governance layer (L3) sits between the agent reasoning layer (L2) and the tool proxy execution layer (L4). There is no path from L2 to L4 that does not pass through L3. The agent produces proposals. The governance layer evaluates them. Only approved proposals reach execution.

Every action proposal crossing the governance admission boundary is validated against the current governance state: schema validation, identity authentication, capability normalization, and threat posture classification all occur before policy evaluation begins. If governance state cannot be established — if the operational mode is undefined, if the threat posture is unclassifiable, if the authority context is missing — the request is denied at the boundary.

Tip

Governance state is not static. It changes when threat level escalates, when authority context changes, and when policy is updated. The governance layer always evaluates against the current state — not the state at session start.

---

Failure Mode

A system that executes before governance evaluates it is a system where governance is, at best, observational. Observational governance can tell you what happened. It cannot prevent what is happening. The failure mode is not a dramatic bypass of governance — it is the gradual normalization of pre-execution action: first in low-risk cases (“this is obviously safe”), then in routine cases (“we always do this”), then as the default. By the time governance is consistently bypassed for anything except the highest-risk actions, it has ceased to be governance and become an exception handler for edge cases. Governance Before Execution is the structural commitment that prevents this normalization from beginning.

Danger

This is constitutionally prohibited. An action that executes before governance evaluates it is not a governed action. It is an ungoverned action that was subsequently logged.

---

Relationship to Constitution

Governance Before Execution is the doctrinal foundation of Article III — Deterministic Enforcement. The constitutional requirement that the governance runtime be positioned between AI agents and all operational infrastructure — with no direct execution path from agent to infrastructure — is the architectural enforcement of this doctrine. It also grounds Article IV — Human Oversight: human oversight is only meaningful if governance evaluation, including the escalation pathway to human review, occurs before execution reaches infrastructure.

---

Footnotes

1. K. Ogata, Modern Control Engineering, 5th ed. Prentice Hall, Upper Saddle River, NJ, 2010. See REFERENCES.md. ↩