GRDL Language Reference

GRDL Language Reference

Complete specification of the GRDL YAML format.

Ruleset (top level)

id: my-ruleset-v1
name: My Governance Rules
description: Governance for my agents
version: "1.0.0"
framework: enterprise_policy    # see framework values below
default_enforcement: enforce
default_severity: high
graceful_degradation: allow_with_audit
rules:
  - ...

Framework values

The framework field documents which governance framework your rules implement. Metadata only — the engine treats all rules identically.

Value	Use case
enterprise_policy	Corporate AI governance
dao_charter	DAO or cooperative bylaws
ai_safety	AI safety frameworks
eu_ai_act	EU AI Act compliance
nist_ai_rmf	NIST AI Risk Management Framework
iso_42001	ISO 42001 AI management
custom	Any other governance document

Graceful degradation policies

Value	Behavior
allow_with_audit	Allow action but log the failure (default)
deny_all	Block all actions until engine recovers
last_known_good	Use cached verdict for this action type

Rule

- id: safety.budget_cap
  name: Budget enforcement
  laws: [SAFETY]
  policy_refs:
    - framework: nist_ai_rmf
      provision_id: "MAP 1.5"
  scope: resource_consumption
  condition:
    field: payload.estimated_cost
    operator: gt
    value_source: context.remaining_budget
  severity: critical
  enforcement: enforce
  target: runtime
  remedy:
    action: block
    message: Cost exceeds budget
    escalation: operator
  timeout_ms: 20

Target (compilation destination)

Targets are runtime-agnostic — each backend maps them to its own enforcement:

Target	What it controls	Example
infrastructure	Filesystem, process, kernel-level protections	Audit log read-only, non-root execution
network	Network endpoint control with protocol inspection	API endpoint restrictions
runtime	CFAIS sidecar evaluation per request	Budget checks, PII controls, identity
hybrid	Both network + runtime	API restriction AND semantic check

Legacy values openshell_static and openshell_dynamic are accepted and auto-normalized to infrastructure and network.

Laws

Value	Law
PRIMACY	Human authority overrides agent decisions
TRANSPARENCY	All decisions must be explainable
ACCOUNTABILITY	Every action traced to an entity
FAIRNESS	No bias in agent decisions
SAFETY	Prevent harm, halt on uncertainty
PRIVACY	Minimum data, consent required
GRACEFUL_DEGRAD	Deterministic fallback on failure

Severity

Value	Engine behavior
critical	Block action, full audit trail
high	Block action, log warning
medium	Allow with audit flag
low	Allow, log for analytics
advisory	Log only

Enforcement

Value	Behavior
enforce	Active blocking (403)
audit	Log violations but allow (200 with audit flag)
shadow	Evaluate silently — for testing new rules

Conditions

Deterministic predicates. Never probabilistic.

Atomic

condition:
  field: payload.amount
  operator: gt
  value_source: context.limit

Operators

Operator	Description	Value type
eq	Equal	any
ne	Not equal	any
gt	Greater than	numeric
gte	Greater or equal	numeric
lt	Less than	numeric
lte	Less or equal	numeric
in	Value in list	list
not_in	Value not in list	list
contains	String contains	string
exists	Field exists. value: true = must exist, value: false = must be absent	boolean
between	Within range	[low, high]

Composite

condition:
  logic: all_of          # AND. Also: any_of (OR), none_of (NOT)
  conditions:
    - field: payload.amount
      operator: gt
      value: 10000
    - field: payload.approved
      operator: eq
      value: false

Composites nest arbitrarily.

Field resolution

Prefix	Source
payload.X	Action payload
context.X	Organizational context (budgets, thresholds)
agent_id	Agent identifier
action_type	Action type
http_method	HTTP method
config.X	Engine configuration

Remedy

remedy:
  action: block
  message: "Why this was blocked"
  alternative: "How to accomplish it compliantly"
  escalation: operator
  audit_log: true

Every denial is transparent and actionable (Law 2).

Policy references

policy_refs:
  - framework: eu_ai_act
    provision_id: article_14
    description: Human oversight for high-risk AI

Metadata for auditors — the engine doesn't interpret these.