Covenants

Governance That Travels With Your Data

The Problem We Solve

When your data moves between systems, applications, and AI tools, how do you ensure your governance policies follow? Traditional approaches fail because:

  • Policy enforcement stops at system boundaries. Your compliance rules live in one system, but your data travels everywhere.
  • Governance metadata gets lost. By the time data reaches its destination, the handling requirements are gone.
  • AI systems have no visibility. Large language models and automation tools don't know what they can and cannot do with your data.
  • Manual compliance is unsustainable. Your teams spend hours coordinating data handling rules across disconnected systems.

Covenant Definition Language (CDL) changes this. Your governance rules travel cryptographically attached to your data, understood and enforced by every AI system that touches it.  Realflow created CDL.  And only we make it this easy to implement.

How CDL Works

CDL attaches three types of declarations directly to your data (via x-extension properties on the JSONSchema definition):

What Your Data IS (Classifications)

Tag data with its sensitivity level and regulatory requirements at any level of detail—from entire records down to individual fields.

  • Sensitivity Levels: Public, Internal, Confidential, Restricted
  • Regulatory Categories: PHI (healthcare), PII (personal data), PCI (payment cards), Financial
  • Jurisdictional Tags: GDPR, CCPA, HIPAA, FERPA compliance indicators

What You MUST or MUST NOT Do (Covenants)

Define binding constraints on how data can be used, stored, and transmitted.

  • Storage Controls: No persistence, no caching, no logging, retention periods
  • Transmission Controls: No external transmission, encryption required, no export
  • AI-Specific Controls: No training on this data, no model persistence
  • Access Controls: Healthcare staff only, role-based restrictions, audit requirements

What Your Tools CAN Do (Capabilities)

Every tool and service declares what it's capable of doing—sending emails, writing to storage, calling external APIs, logging operations.

The magic: LLM systems automatically match data covenants against tool capabilities. If there's a conflict, the operation is blocked before it happens.

Key Capabilities

Field-Level Governance

Apply different rules to different parts of your data. A patient record might allow the name to be shared internally, but restrict diagnosis codes to physicians only.  Grades might be protected by FERPA while a student's vaccination status carries separate HIPAA requirements.  These classifications can be applied at any level; to the entire data source, to specific entities (tables), or to specific entity attributes (columns)

Covenants inherit through your data hierarchy. Set rules at the top level; they automatically apply to everything beneath unless you explicitly override them for specific fields.

Human-in-the-Loop Integration

When governance violations are detected, the system doesn't just block-it engages your people:

  • Request human approval for exceptions
  • Suggest compliant alternatives
  • Generate secure links for authorized access
  • Create redacted versions that satisfy constraints

These approval workflows can pause for days or weeks, send notification links via SMS or email to external stakeholders, and resume automatically when approvals arrive.

Operator Visibility Controls

A breakthrough capability: AI can process sensitive data and provide valuable conclusions to operators who aren't authorized to see the raw data.

Example: A physician assistant receives guidance about a patient's condition and recommended next steps, while actual lab values and diagnosis codes remain hidden; processed by the AI but filtered from display.

  • No-display-to-operator: AI processes data internally but filters sensitive fields from output
  • Redacted display: Show partial information (last 4 digits, masked values)
  • Summary-only display: Provide conclusions without raw data
  • Role-restricted display: Different visibility based on user role

Cryptographic Trust

CDL declarations can be cryptographically signed, proving:

  • Authenticity: Declarations came from an authorized source
  • Integrity: Declarations haven't been tampered with
  • Validity Period: Declarations have a defined lifetime

Governance requirements can't be forged or bypassed-they can travel with the data as verifiable, tamper-proof attestations.

Semantic Intelligence

Unlike rigid rule engines, CDL leverages an LLM's natural language understanding. The system comprehends that:

  • "logs-operations" implies data persistence
  • "can-send-email" conflicts with "no-external-transmission"
  • "audit-required" is satisfied by "logs-operations"

You declare intent in clear language; the LLM understands implications and relationships.

Common Use Cases

Healthcare: HIPAA-Compliant AI Workflows

A patient record carries PHI classification with HIPAA covenants. When processed by an LLM:

  • The LLM understands what it can and cannot do
  • Logging tools that would persist PHI are automatically excluded
  • Email notifications are blocked for sensitive data
  • Role-based display ensures only authorized staff see raw clinical data
  • Full audit trail maintained for compliance

Financial Services: PCI and Privacy Protection

Payment card data carries PCI covenants that prevent:

  • Storage in non-compliant systems
  • Transmission to external APIs
  • Display to operators without proper authorization
  • Caching in any form

The rules travel with the card data through every system that processes it.

Multi-Party Workflows

Share data with partners while maintaining control:

  • Attach covenants before data leaves your boundary
  • Partner systems automatically enforce your requirements
  • Violations are blocked at the source
  • Audit trails prove compliance across organizational boundaries

LLM Agent Governance

Autonomous LLM agents that orchestrate multiple tools:

  • Agent manifests declare aggregated capabilities
  • Data covenants prevent agents from routing sensitive data to inappropriate tools
  • Human-in-the-loop approvals for high-risk operations
  • Expiring links ensure time-bounded access

Integration Points

CDL works across your entire stack:

  • MCP Tool and Agent Declarations: Individual tools declare their capabilities
  • Server-Level Defaults: Set capabilities for all tools on a server
  • API Gateway Injection: Add governance to legacy systems through Azure APIM, AWS API Gateway, or Kong
  • Response-Time Declaration: Capture what actually happened during execution
  • Separate Governance Endpoints: Centralized compliance team control via sidecar pattern

The Realflow.ai Advantage

CDL is an open specification—we're releasing it for the industry to adopt. But Realflow.ai SaaS provides:

  • Visual Workflow Builder: Create CDL-governed workflows without coding
  • Pre-Built Connectors: 185+ integrations with CDL starter kits built-in
  • Hosted Governance Platform: Managed enforcement, audit trails, and reporting
  • Human-in-the-Loop: Native support for pause/resume approval workflows
  • Enterprise Security: Double envelope encryption, role-based access

Your data governance rules should travel with your data. CDL makes it possible. Realflow.ai makes it easy.