GRDL — Governance Rule Definition Language

The YAML standard for AI agent governance. Runtime-agnostic. Deterministic. Sub-microsecond.

GRDL compiles structured governance rules into any enforcement backend — NVIDIA OpenShell, Docker, Kubernetes, or standalone HTTP. One ruleset, any runtime.

---

The problem

Autonomous AI agents (Gemma 4, OpenClaw, Claude Code, custom LLM agents) make decisions and take actions without human input. The industry has infrastructure security (OpenShell, Docker sandboxes). Nobody has decision governance — should this agent spend beyond its budget? Can it escalate its own permissions? Is the delegation chain too deep?

GRDL is the missing layer.

How it works

Agent decides to act → CFAIS sidecar evaluates → ALLOW or DENY → Agent proceeds or adjusts

Write governance rules in YAML. The compiler produces enforcement config for your runtime:

prahari compile rules.grdl.yaml --backend openshell    # NVIDIA OpenShell YAML
prahari compile rules.grdl.yaml --backend docker       # Docker Compose + seccomp
prahari compile rules.grdl.yaml --backend standalone   # HTTP sidecar only (works anywhere)

The CFAIS engine is deterministic — mathematical comparisons, not ML. Sub-microsecond evaluation. Single static binary.

Quick start

go install github.com/prahari-ai/grdl/cmd/prahari@latest

# Pick a template, compile, test
prahari validate examples/templates/enterprise-agent-governance.grdl.yaml
prahari compile examples/templates/enterprise-agent-governance.grdl.yaml --backend standalone
prahari evaluate examples/templates/enterprise-agent-governance.grdl.yaml test-action.json

# Run the governance sidecar
prahari serve examples/templates/enterprise-agent-governance.grdl.yaml --addr :9700

Any agent framework (Gemma 4 + Ollama, OpenClaw, LangChain, CrewAI) calls POST http://localhost:9700/evaluate before executing each tool call. Denied actions return 403 with an explanation.

Backends

Backend	Infrastructure enforcement	Use case
openshell	OpenShell YAML (filesystem, network, process, Landlock)	NVIDIA NemoClaw sandboxes
docker	Docker Compose + seccomp profiles	Standard container deployments
standalone	None (advisory HTTP sidecar)	Any agent, any runtime, any language

Same GRDL rules, same CFAIS engine, different infrastructure layer.

Governance templates

Template	Rules	Key governance controls
enterprise-agent-governance	11	Budget caps, privilege escalation, PII controls, cascading depth, human-in-the-loop
dao-governance	4	Treasury limits, quorum, voting weight caps
ai-safety	4	Tool allowlists, output validation, rate limiting

The Seven Constitutional Laws

GRDL encodes seven constitutional laws for Autonomous Intelligent Systems (CFAIS). Each law compiles to an enforceable GRDL rule.

#	Law	Constraint	GRDL enforcement
1	Human Dignity	Actions must not violate human dignity as defined by the applicable constitutional framework.	VERIFY action.impact_on_humans.dignity_score >= constitutional_threshold
2	Transparency	The system provides a complete explanation of its reasoning on request by authorized parties.	REQUIRE action.explanation.completeness >= 0.95 AND explanation.constitutional_trace = true
3	Accountability	The system maintains an immutable audit trail of all decisions and actions.	LOG decision_chain TO accountability_layer WITH cryptographic_integrity = true
4	Human Override	The system accepts override commands from authorized parties, except where the override would itself cause a constitutional violation.	ACCEPT human_command IF authorized(commander) AND NOT violates(command, constitutional_constraints)
5	Self-Limitation	The system does not acquire capabilities beyond the scope necessary for its assigned function.	CONSTRAIN capability_acquisition TO function_scope AND ALERT IF detected_capability > authorized_capability
6	Value Alignment	The system's actions remain aligned with encoded human constitutional values.	VERIFY action.value_alignment_score >= alignment_threshold
7	Graceful Degradation	On constraint violation or uncertainty, the system degrades to a verified safe state rather than failing open.	ON violation OR uncertainty: DEGRADE TO nearest_verified_safe_state

Governance parameters

Laws operate over configurable governance parameters. Proportionality bounds the influence of autonomous agents relative to humans:

# Proportionality: AI voting weight is bounded to a fraction of human-equivalent weight
governance_parameters:
  proportionality:
    ai_weight_bound: 0.25        # alpha: max AI weight as fraction of human-equivalent
    scope: [dao_voting, delegation]
    enforcement: REQUIRE ai_agent.weight <= alpha * human_equivalent.weight

Proportionality is a governance parameter enforced under Law 4 (Human Override) and Law 5 (Self-Limitation), not a separate law.

License

Core: Apache 2.0 | Enterprise modules: BSL 1.1

Built by Prahari.ai