Compliance Framework

Control mappings, audit procedures, and regulatory alignment

Executive Summary

RayRay is architected for deployment in controlled unclassified environments (CUI) with alignment to FedRAMP authorization boundaries. The system implements a compliance-first approach where AI-assisted operations maintain human oversight as mandated by OMB M-25-21 ("Advancing Governance, Innovation, and Risk Management for Agency Use of AI").

Applicable Frameworks

Framework	Authority	Applicability
OMB M-25-21	Office of Management and Budget	Mandatory for Federal AI systems
NIST 800-53 Rev 5	NIST	Security and privacy controls
FedRAMP	GSA	Cloud service authorization
FISMA	DHS	Information security management
NIST AI RMF	NIST	AI risk management

Control Family Mappings

Access Control (AC)

Control	Implementation	Evidence
`AC-2` Account Management	Role-based accounts with email-based identification	User audit table, role assignments
`AC-3` Access Enforcement	Four-tier role model (Observer, Analyst, Reviewer, Admin)	JWT role claims, endpoint guards
`AC-6` Least Privilege	Role permissions enforced at API layer	Permission matrix documentation

Identification and Authentication (IA)

Control	Implementation	Evidence
`IA-2` Identification and Authentication	OAuth 2.0 password grant with JWT tokens	Auth service logs, token issuance records
`IA-2(1)` Multi-Factor Authentication	TOTP-based MFA (RFC 6238 compliant)	MFA enrollment logs, challenge records
`IA-5` Authenticator Management	bcrypt password hashing, TOTP secrets encrypted	Password policy enforcement logs

Audit and Accountability (AU)

Control	Implementation	Evidence
`AU-2` Audit Events	Comprehensive event logging for all CUD operations	Event type enumeration in codebase
`AU-3` Content of Audit Records	Timestamp, user ID, IP, resource type, action, outcome	JSON Lines format specification
`AU-4` Audit Storage Capacity	Daily log rotation with configurable retention	Storage configuration documentation
`AU-9` Protection of Audit Information	SHA-256 hash chain prevents tampering	Hash verification API endpoint
`AU-11` Audit Record Retention	Configurable retention period (default: 7 years)	Retention policy configuration

OMB M-25-21 Alignment

OMB M-25-21 mandates that agencies ensure human oversight of AI systems that could impact rights or safety. RayRay implements the following safeguards:

Required: Human-in-the-Loop (HITL)

┌─────────────────┐
│   AI Extraction │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Redundancy Check│  ◀── Quality validation
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  CHECKPOINT     │  ◀── M-25-21 MANDATORY
│  Human Review   │      Human must approve
└────────┬────────┘      before ANY commit
         │
    ┌────┴────┐
    ▼         ▼
┌───────┐ ┌───────┐
│Commit │ │Reject │
└───────┘ └───────┘

Key compliance point: The LangGraph workflow enforces a hard checkpoint at the human review stage. No database write occurs without explicit human action. This satisfies M-25-21's requirement for "meaningful human oversight" of AI outputs.

AI Transparency Requirements

Requirement	Implementation
AI-generated content identification	All extractions marked with AI model, provider, and confidence score
Human modification tracking	Original AI values preserved when human edits occur
Review audit trail	Reviewer ID, timestamp, and decision recorded for every commit

Audit Log Architecture

Hash Chain Integrity

Each audit event includes a SHA-256 hash computed over the event content plus the previous event's hash. This creates a blockchain-style integrity chain where tampering with any historical event breaks the chain.

// Event structure
{
  "_id": "evt_abc123",
  "_timestamp": "2026-03-12T15:30:00Z",
  "_prev_hash": "sha256:7f8a...",
  "_hash": "sha256:3b2c...",
  "event_type": "review_approve",
  "user_id": 42,
  "resource_type": "extraction",
  "resource_id": 123,
  "outcome": "success"
}

// Integrity verification
GET /api/audit/verify
{
  "valid": true,
  "events_checked": 1847,
  "first_failure": null,
  "errors": []
}

Event Types

Event Type	Description	NIST Control
`document_upload`	Document submitted for processing	AU-2
`extraction_start`	AI extraction initiated	AU-2
`extraction_complete`	AI extraction finished with confidence scores	AU-2
`review_approve`	Human approved AI extractions	AU-2, M-25-21
`review_modify`	Human modified AI extractions	AU-2, M-25-21
`review_reject`	Human rejected AI extractions	AU-2, M-25-21
`export_download`	Report exported (PPTX, PDF)	AU-2

Role-Based Access Control

RayRay implements a four-tier role model aligned with separation of duties principles:

Role	Permissions	Typical Assignment
Observer	Read-only access to documents and extractions	Stakeholders, Auditors
Analyst	Upload documents, review drafts, mark for approval	Data entry staff, Junior analysts
Reviewer	Approve/reject extractions, commit to database	Senior analysts, Supervisors
Admin	Full system access, user management, configuration	System administrators, ISSOs

Separation of Duties

M-25-21 Control: The system enforces that analysts cannot commit AI extractions to the database — only reviewers and admins hold commit authority. This prevents single points of failure and ensures appropriate oversight.

// Permission matrix
┌─────────────┬─────────┬─────────┬──────────┬────────┐
│  Action     │Observer │ Analyst │ Reviewer │ Admin  │
├─────────────┼─────────┼─────────┼──────────┼────────┤
│ View        │    ✓    │    ✓    │    ✓     │   ✓    │
│ Upload      │    ✗    │    ✓    │    ✓     │   ✓    │
│ Edit Draft  │    ✗    │    ✓    │    ✓     │   ✓    │
│ Commit      │    ✗    │    ✗    │    ✓     │   ✓    │
│ User Mgmt   │    ✗    │    ✗    │    ✗     │   ✓    │
└─────────────┴─────────┴─────────┴──────────┴────────┘