AEGIS System Principles

Architectural Enforcement & Governance of Intelligent Systems

Version: 0.2
Status: Informational
Part of: AEGIS Architecture
Author: Kenneth Tannenbaum
Last Updated: March 6, 2026

Purpose

These principles are normative architecture rules for AEGIS. They define what must remain true for the system to be secure, governable, and auditable.

P1: Bounded Capability

Rule:

• No actor may exercise capability outside explicitly authorized scope.

Implementation check:

• Capability + resource + scope must be validated before execution.

P2: Complete Mediation

Rule:

• Every capability invocation must pass governance evaluation.¹²

Implementation check:

• No direct execution path from agent layer to infrastructure.

P3: Default Deny

Rule:

• Absence of explicit authorization is treated as denial.²

Implementation check:

• No policy match results in DENY.

P4: Deterministic Governance

Rule:

• Same input + policy version + context yields same decision.³

Implementation check:

• Replay test corpus has zero decision mismatches.

P5: Auditability

Rule:

• Every decision and execution must have immutable evidence.

Implementation check:

• Execution event without prior audit ID is treated as policy violation.

P6: Explicit Authority Boundaries

Rule:

• Governance authority and execution authority are separate.

Implementation check:

• Decision Engine authorizes; Tool Proxy executes; neither role is conflated.

P7: Fail-Closed Safety

Rule:

• System uncertainty or subsystem failure must not produce implicit allow.³

Implementation check:

• Error paths return ESCALATE or DENY.

P8: Least Privilege by Construction

Rule:

• Runtime permissions are narrowed to minimum required scope and duration.⁴²

Implementation check:

• Constrained decisions carry enforceable limits (time, rate, target, size).

P9: Policy Integrity

Rule:

• Authorization logic depends only on authenticated policy artifacts.

Implementation check:

• Unsigned or modified policy bundles are rejected.

P10: Human Accountability for Exceptions

Rule:

• Break-glass or escalated paths require accountable human oversight.

Implementation check:

• Escalation approvals are attributable and fully audited.

Standards Alignment

These principles collectively implement the GOVERN and MANAGE functions of the NIST AI Risk Management Framework (AI RMF 1.0),⁵ the risk management, record-keeping, and quality management requirements of the EU Artificial Intelligence Act,⁶ and the operational control and monitoring obligations of ISO/IEC 42001:2023.⁷

Principle Compliance Review

Each release SHOULD include a principle compliance checklist proving:

• No mediation bypass.
• Deterministic replay parity.
• Immutable audit linkage.
• Fail-closed behavior under injected failures.

References