AEGIS Policy Language

Architectural Enforcement & Governance of Intelligent Systems

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

Overview

Policies are expressed in a structured YAML‑based DSL that defines authorization rules for capability requests.¹

Policy Structure

policy_id: filesystem_read_policy
name: "Public Data Read Access"
effect: allow
priority: 100
enabled: true
conditions:
  - type: capability
    value: filesystem.read
  - type: resource_prefix
    value: /data/public
  - type: actor_role
    value: system_agent
constraints:
  max_size_mb: 100
  rate_limit: "10/minute"
risk_modifier: -10

Fields

Effects

Condition Types

Capability Match

- type: capability
  value: filesystem.read
  # Matches exact capability or parent
  # filesystem.read matches: filesystem, filesystem.read (not filesystem.write)

Resource Patterns

- type: resource_prefix
  value: /data/public
  # Prefix match: matches /data/public/*, /data/public**

- type: resource_exact
  value: /etc/passwd
  # Exact match only

- type: resource_regex
  value: "^/var/log/.*\.log$"
  # Regex pattern match

Actor Conditions

- type: actor_id
  value: agent_123
  # Exact agent ID match

- type: actor_role
  value: system_agent
  # Actor must have role

- type: actor_trust
  comparison: ">"
  value: 80
  # Actor trust score threshold

Environment Conditions

- type: environment
  value: production
  # Environment must match

- type: time_window
  start: "09:00"
  end: "17:00"
  # Request within business hours

- type: day_of_week
  values: [Mon, Tue, Wed, Thu, Fri]
  # Weekdays only

Constraint Types

Applied when effect=constrain:

constraints:
  max_size_mb: 100              # File size limit
  max_rows: 1000               # Query result limit
  rate_limit: "10/minute"      # Requests per time
  timeout_seconds: 30          # Execution timeout
  audit_required: true         # Force full audit
  log_level: INFO              # Logging verbosity

Example Policies

Policy 1: Allow Public Data Read

policy_id: allow_public_read
name: "Allow reading public data"
effect: allow
priority: 100
conditions:
  - type: capability
    value: filesystem.read
  - type: resource_prefix
    value: /data/public
risk_modifier: -5

Policy 2: Block Sensitive Files

policy_id: block_sensitive_files
name: "Block access to sensitive system files"
effect: deny
priority: 10
conditions:
  - type: capability
    value: filesystem.read
  - type: resource_regex
    value: "^(/etc/shadow|/etc/passwd|/etc/sudoers)$"

Policy 3: Constrained Database Access

policy_id: db_query_constrained
name: "Constrained database queries"
effect: constrain
priority: 50
conditions:
  - type: capability
    value: data.database_query
  - type: environment
    value: production
constraints:
  max_rows: 5000
  rate_limit: "5/minute"
  timeout_seconds: 30
  audit_required: true
risk_modifier: 5

Policy 4: Escalate Critical Operations

policy_id: escalate_critical_ops
name: "Escalate critical system operations"
effect: escalate
priority: 5
conditions:
  - type: capability
    value: compute.process_spawn
  - type: actor_role
    value: ai_agent

Policy 5: Time-Windowed Access

policy_id: business_hours_only
name: "Restrict sensitive operations to business hours"
effect: deny
priority: 20
conditions:
  - type: capability
    value: data.api_call
  - type: resource_prefix
    value: https://sensitive-api.internal
  - type: time_window
    start: "17:00"
    end: "09:00"
    # Denies outside 9am-5pm

Condition Evaluation

All conditions within a policy are AND-ed:

Policy matches if:
  (capability == filesystem.read)
  AND (resource matches /data/public)
  AND (actor_role == system_agent)

Policies with multiple conditions are more specific and prefer EXACT/REGEX matches over PREFIX matches.

Conflict Resolution

When multiple policies match:

1. Explicit DENY always wins (if any DENY matches)²
2. Priority field determines evaluation order (higher first)
3. First matching policy wins (for ALLOW/CONSTRAIN/ESCALATE)
4. Default-deny if no policies match²

Performance Considerations

• Policy evaluation is O(n) where n = number of policies
• Regex conditions are slower than prefix/exact
• Sort policies by specificity + priority for optimal performance
• Cache compiled regex patterns

References