Spec 12: Compliance & Governance

DWS Spec 12: Compliance & Regulatory Governance

Digital Worker Standard — DWS Specification

Version: 1.0 Tier: 4 — Governance Status: Release Candidate Dependencies: Spec 0 (Project Structure), Spec 1 (Worker Identity), Spec 11 (Events & Telemetry)

1. Overview

This specification defines how DWS definitions serve as compliance artifacts. A job spec is not just a technical specification for a digital worker; it is also a governance document that demonstrates how the worker satisfies regulatory requirements.

Regulatory frameworks increasingly require organisations to demonstrate structured human oversight, traceability, risk management, and accountability for AI systems. DWS’s existing primitives (authority levels, verification gates, event streams, approval gates) already satisfy many of these requirements. This spec formalises the mapping and adds the metadata needed to make compliance auditable.

2. Risk Classification

Every job spec MAY declare a risk classification in its manifest (Spec 0, Section 2.1). The classification determines which governance requirements apply.

2.1 Risk Tiers

Tier	Label	Description	Governance Requirements
0	`minimal`	Low-risk worker with no significant impact.	Event logging only.
1	`limited`	Worker produces output that is visible to end users.	Transparency requirements. Must disclose AI involvement.
2	`high`	Worker operates in regulated domains or makes decisions affecting people.	Full governance: human oversight, verification gates, approval gates, audit retention.
3	`unacceptable`	Prohibited use cases.	DWS MUST NOT be used for prohibited purposes. Runtimes MUST refuse to instantiate.

These tiers align with the EU AI Act risk classification. Organisations subject to different frameworks (NIST RMF, ISO 42001, Australian AI governance) can map their framework requirements to these tiers via the frameworks field.

2.2 Schema

{
  "compliance": {
    "risk_classification": "high",
    "frameworks": ["eu-ai-act", "nist-rmf", "iso-42001"],
    "human_oversight_required": true,
    "audit_retention_days": 365,
    "transparency": {
      "disclose_ai_involvement": true,
      "disclosure_method": "output_metadata"
    }
  }
}

3. Regulatory Mapping

DWS primitives map to regulatory requirements:

3.1 EU AI Act (August 2026)

EU AI Act Requirement	DWS Primitive	Evidence
Human oversight	Authority levels (Spec 1), Approval gates (Spec 10)	Worker descriptor shows authority level. Workflow shows approval gate placement.
Risk management	Risk classification (this spec), Guardrails (Spec 1)	Manifest declares risk tier. Guardrails document safety measures.
Record-keeping	Event stream (Spec 11)	Append-only event store with retention policies.
Transparency	Transparency block (this spec)	Disclosure method documented. Outcome envelopes include worker identity.
Technical documentation	Worker descriptor + full job spec	The job spec IS the technical documentation.
Accuracy, robustness	Verification framework (Spec 8)	Verification gates with structured findings provide ongoing accuracy measurement.
Corrective actions	Re-verification cycle (Spec 8), Escalation (Spec 7)	Failed verification triggers rework. Exhausted retries trigger escalation to humans.

3.2 NIST AI Risk Management Framework

NIST Function	DWS Primitive
Govern	Risk classification, compliance block, authority levels
Map	Worker descriptor (domain, role, boundaries)
Measure	Verification framework, event telemetry, cost tracking
Manage	Escalation protocol, approval gates, knowledge lifecycle

3.3 ISO/IEC 42001

DWS definitions provide the technical implementation of ISO 42001 management system requirements. The job spec artifacts (git-committed, versioned, reviewed) serve as the documented AI policies that ISO 42001 requires.

4. Audit Trail Requirements

4.1 Minimum Audit Events

For workers classified as high risk, the following events MUST be retained:

All lifecycle.* events
All workflow.* events
All verification.* events
All approval.* events
All intent.* events
All guardrail.blocked events
All coordination.escalation_* events

4.2 Retention

Risk Tier	Minimum Retention	Recommended
`minimal`	30 days	90 days
`limited`	90 days	1 year
`high`	1 year	5 years

Retention periods are measured from event timestamp. Runtimes MUST NOT delete events before their retention period expires. Organisations MAY configure longer retention.

4.3 Audit Export

Runtimes MUST support exporting audit-relevant events in a structured format suitable for regulatory review. The export MUST include:

Event data with all common fields (Spec 11, Section 2)
Worker descriptor versions active during the audit period
Intent artifacts (both git-committed and runtime mutations)
Verification verdicts and findings

5. Compliance Validation

The dws validate tool SHOULD check compliance configuration:

Errors (block deployment):

Worker classified as high with no verification gates in any workflow.
Worker classified as high with authority: autonomous and no approval gates.
Worker classified as high with audit_retention_days below the minimum for its frameworks.

Warnings:

Worker with no risk classification declared.
Worker with human_oversight_required: true but no supervised or collaborative interaction modes in any workflow phase.
Worker handling confidential or restricted data with no guardrails.

6. OWASP Agentic Risk Mapping

The OWASP Top 10 for Agentic Applications (2026) identifies risks that DWS addresses:

OWASP Risk	DWS Mitigation
Excessive agent authority	Authority levels (Spec 1) with graduated trust. Boundary declarations prevent scope creep.
Unsafe tool use	Tool bindings in skills (Spec 3) with required/optional flags. Guardrails on tool inputs/outputs.
Broken access control in multi-agent systems	Coordination protocol (Spec 7) with awareness boundaries. Worker-to-worker contracts.
Knowledge poisoning	Knowledge entry confidence scoring, decay model, and data classification (Spec 2).
Insufficient output verification	Verification framework (Spec 8) with independent, context-isolated verifiers.
Lack of audit trail	Event stream (Spec 11) with mandatory events and retention policies.

7. References

EU AI Act — Risk classification framework and regulatory requirements.
NIST AI Risk Management Framework — Federal AI governance framework.
ISO/IEC 42001 — International standard for AI management systems.
OWASP Top 10 for Agentic Applications (2026) — Security risk taxonomy for AI agents.
Spec 0: Project Structure — Manifest compliance block.
Spec 1: Worker Identity — Authority levels and guardrails.
Spec 8: Verification Framework — Quality evaluation and evidence.
Spec 11: Events & Telemetry — Event retention and audit trail.