Human Oversight Policy & Governance

Relevant source files

The following files were used as context for generating this wiki page:

• governance/human-oversight/governance-board.md
• legal/HUMAN-OVERSIGHT-POLICY.md
• src/lib/governance/human-approval.ts
• governance/README.md
• governance/system-cards/isabella-5-3-0.md
• server.ts

The Human Oversight Policy and Governance framework establishes a native control layer within Isabella to ensure human authority over high-impact AI actions. This framework aligns with international standards from UNESCO, the UN, and the WEF to prioritize human rights, transparency, and accountability.

The system implements oversight as a runtime contract (HumanApproval) rather than a simple documentation layer. It enforces tiered approval levels (H0-H4) for specific operations, preventing autonomous execution of actions classified as high-risk or prohibited without explicit human verification.

Sources: governance/human-oversight/governance-board.md:1-12, governance/README.md:3-6, legal/HUMAN-OVERSIGHT-POLICY.md:3-6

Oversight Levels (H0 - H4)

The project categorizes all system actions into five distinct levels of oversight. These levels determine whether an action proceeds automatically or requires intervention from one or more authorized reviewers.

Level	Name	Rule	Action Result
H0	No Impact	Automation permitted	Immediate execution
H1	Post-Review	Sampled review after execution	Execution + log for audit
H2	Human Approval	Approval required before execution	pending_human_approval
H3	Double Approval	Two independent approvers required	Blocked until dual sign-off
H4	Human Exclusive	Autonomous execution prohibited	Blocked (403)

Sources: src/lib/governance/human-approval.ts:11-20, legal/HUMAN-OVERSIGHT-POLICY.md:8-14, governance/system-cards/isabella-5-3-0.md:34-40

The Governance Board

The Governance Board serves as the institutional body responsible for overseeing model deployments, policies, and risk management.

Key Responsibilities

• Approving model deployments and tool policies via H3 (Double Approval).
• Maintaining and updating the AI Risk Register.
• Reviewing and closing SEV-1 and SEV-2 incidents.
• Resolving conflicts between risk owners and curators.

Sources: governance/human-oversight/governance-board.md:14-20, governance/README.md:14-20

Technical Implementation: Human Approval Gate

The human-approval.ts module provides the runtime logic for enforcing oversight. It includes data structures to track approval states and a middleware for Express routes to intercept high-impact requests.

Data Structures

The HumanApproval interface defines the requirements for a valid decision:

• approvalId: A unique identifier for the approval record.
• reviewerSubject & reviewerRole: Identifies the person with authority making the decision.
• evidenceRefs: Links to technical evidence justifying the risk level.
• expiresAt: The timestamp when the approval becomes invalid.

Sources: src/lib/governance/human-approval.ts:22-35, legal/HUMAN-OVERSIGHT-POLICY.md:16-21

Logic Flow and Anti-Evasion

The system enforces strict anti-evasion rules:

1. No Anonymous Approval: reviewerSubject is mandatory.
2. Mandatory Reasoning: Approvals without a provided reason are invalid.
3. Independence: For H3 levels, the second reviewer cannot be the same as the first.
4. Expiration: Approvals have a TTL; expired approvals trigger a re-request for oversight.

Integration with Server API

The server integrates the humanApprovalGate middleware into routes deemed "High Impact." When a request hits a governed endpoint without proper human sign-off, the server responds with specific status codes.

Governance API Endpoints

• GET /api/v1/governance/risk-register: Retrieves the current list of AI risks and their tiers.
• GET /api/v1/governance/readiness: Admin endpoint to check if the system meets production deployment requirements.
• GET /api/v1/governance/provenance: Returns metadata regarding the origin and review status of system data.

Sources: server.ts:397-425, governance/README.md:22-26

High-Impact Actions

Actions requiring H2 or higher typically include:

• Deployment of new AI models.
• Financial liquidations or payouts.
• Modification of core system policies.
• Execution of tools in production environments.

Sources: legal/HUMAN-OVERSIGHT-POLICY.md:23-28, governance/system-cards/isabella-5-3-0.md:34-40

Accountability and Artifacts

Governance is supported by versioned artifacts stored in the governance/ directory.

Artifact Type	Path	Purpose
Risk Register	governance/risk-register/	Versioned YAML files tracking AI-RISK identifiers.
Change Records	governance/change-records/	Records of infrastructure and policy mutations (e.g., CR-2026-0001).
System Cards	governance/system-cards/	Architectural and safety documentation for specific versions.
Approvals	governance/approvals/	Audit logs of human decisions made via the approval gate.

Sources: governance/README.md:28-34, governance/change-records/CR-2026-0001.yaml:1-15

Conclusion

The Human Oversight Policy & Governance module ensures that Isabella operates within safe boundaries by embedding human authority directly into the code. By leveraging the tiered H0-H4 system and the native HumanApproval contract, the project maintains transparency and accountability for all high-impact AI decisions.