AEGIS Governed Capability Flow

Architectural Enforcement & Governance of Intelligent Systems

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

Purpose

This document defines the end-to-end capability flow and the control contract for each stage from request proposal to execution outcome.

End-to-End Flow

1) User/System Intent
2) Agent Action Proposal
3) Capability Request Construction
4) Governance Admission Validation
5) Policy Matching and Precedence
6) Risk Calculation
7) Decision Outcome
8) Constraint Packaging (if required)
9) Audit Record Creation
10) Tool Proxy Execution or Block
11) Execution Telemetry and Post-Decision Audit

Stage Contracts

Stage 1-3: Proposal and Request Creation

Required outputs:

• Valid agent_id, capability, resource, context.
• Stable request_id for traceability.

Stage 4: Admission Validation

Required checks:

• Schema validity.
• Identity authenticity.
• Capability format normalization.

Failure behavior:

• Immediate DENY and audit event.

Stage 5: Policy Matching

Required behavior:

• Evaluate enabled policies only.
• Apply deterministic precedence (deny first, then priority).¹

Stage 6: Risk Calculation

Required behavior:

• Compute bounded score in range 0-100.
• Include actor, capability, resource, environment, history factors.

Stage 7-8: Decision and Constraints

Possible outcomes:

• ALLOW
• CONSTRAIN
• ESCALATE
• DENY

Constraint-bearing outcomes must include enforceable machine-readable limits.

Stage 9-11: Audit and Execution

Required behavior:

• Record immutable decision evidence before execution handoff.
• Execute only if decision permits.
• Record execution telemetry and violations.

Flow Invariants

1. No request executes without a prior decision.²
2. No decision exists without an audit ID.²
3. No constrained request executes unconstrained.
4. No denied request causes infrastructure side effects.

Decision Matrix

Verification Checks

The flow is valid only if:

• 100% execution events map to governance decisions.
• All constrained executions show matching runtime controls.
• Escalation queue contains all risk threshold exceedances.
• Replay of flow events is deterministic.

Related Documents

• docs/architecture/DECISION_ALGORITHM.md
• docs/architecture/POLICY_MATCHING_AND_DEBUG.md
• docs/architecture/END_TO_END_REQUEST_FLOW.md

References