MOSS Edge Classifier

On-device policy evaluation and audit logging for AI agent governance.

Overview

The Edge Classifier enables offline-first AI governance by:

• Downloading signed policy bundles from MOSS
• Evaluating policies locally on user devices
• Classifying data (PII detection, keyword matching)
• Buffering audit events and syncing to MOSS
• Operating offline with a configurable grace period
• Verifying ML-DSA-44 post-quantum signatures

Installation

pip install moss-edge-classifier

Quick Start

from moss_edge_classifier import EdgeClassifier

# Initialize with your customer API key
classifier = EdgeClassifier(
    api_key="cust_key_xyz...",
    base_url="https://api.mosscomputing.com",
    offline_grace_period_hours=24,
)

# Download and verify policy bundle
await classifier.refresh_bundle()

# Evaluate an action
decision = await classifier.evaluate(
    agent_id="agent_123",
    action_type="chat",
    input_text="What is the user's SSN?",
    output_text="I cannot share SSN information.",
    context={"user_id": "user_456"},
)

print(decision.allowed)  # True/False
print(decision.policies_evaluated)  # ['pol_pii_block', 'pol_chat_allow']
print(decision.classifications)  # {'input': ['pii_detected'], 'output': []}

Features

Policy Evaluation

• Downloads cryptographically-signed policy bundles
• Evaluates rules locally (no network latency)
• Handles complex policy conditions (capabilities, data classifications)

Data Classification

• PII Detection: SSN, credit cards, email, phone numbers (regex-based)
• Keyword Matching: Custom sensitive terms per organization
• Extensible classification framework

Offline Mode

• Continues working when MOSS API is unreachable
• Configurable grace period (default: 24 hours)
• Automatic bundle refresh on reconnection

Audit Logging

• Buffers audit events locally
• Syncs to MOSS in batches (max 100 per request)
• Idempotent sync (duplicate detection via envelope_id)
• Local ML-DSA-44 signatures on audit entries

Post-Quantum Security

• Verifies bundle signatures using ML-DSA-44
• Downloads MOSS public keys from /v1/keys/signing
• Rejects tampered or unsigned bundles

API Reference

EdgeClassifier

class EdgeClassifier:
    def __init__(
        self,
        api_key: str,
        base_url: str = "https://api.mosscomputing.com",
        offline_grace_period_hours: int = 24,
        max_audit_buffer: int = 1000,
    ):
        """Initialize Edge Classifier.

        Args:
            api_key: Customer API key (starts with cust_key_)
            base_url: MOSS API base URL
            offline_grace_period_hours: Hours to operate offline
            max_audit_buffer: Max audit events before forced sync
        """

Methods

refresh_bundle() Downloads the latest policy bundle from MOSS and verifies the signature.

await classifier.refresh_bundle()

evaluate() Evaluates an action against loaded policies.

decision = await classifier.evaluate(
    agent_id="agent_123",
    action_type="chat",
    input_text="...",
    output_text="...",
    context={},
)

Returns EvaluationDecision with:

• allowed: bool - Whether action is permitted
• decision: str - "allow", "block", or "require_approval"
• policies_evaluated: list[str] - Policy IDs checked
• classifications: dict - Input/output classifications
• bundle_version: int - Policy bundle version used

sync_audit() Syncs buffered audit events to MOSS.

result = await classifier.sync_audit()
# Returns: {"synced": ["env_001", ...], "failed": [], "chainPosition": 42}

get_revocations() Fetches emergency revocation list from MOSS.

revocations = await classifier.get_revocations()
# Returns: {"revocations": [...], "revocationListId": "rvk_xyz"}

Architecture

┌─────────────────────────────────────────────┐
│ AI Agent Application                         │
│  ┌────────────────────────────────────────┐ │
│  │ EdgeClassifier                          │ │
│  │  • Policy Bundle (cached locally)       │ │
│  │  • Signing Keys (cached)                │ │
│  │  • Audit Buffer (SQLite)                │ │
│  │  • Classification Rules                 │ │
│  └────────────────────────────────────────┘ │
│           ▲                      │           │
│           │ Download bundle      │ Sync audit│
│           │ (signed)             │ (batched) │
└───────────┼──────────────────────┼───────────┘
            │                      │
            │    HTTPS (TLS 1.3)   │
            │                      ▼
┌───────────┴──────────────────────────────────┐
│ MOSS API (https://api.mosscomputing.com)     │
│  • GET /v1/customer/policy-bundles/latest    │
│  • GET /v1/keys/signing                      │
│  • POST /v1/customer/audit/sync              │
│  • GET /v1/customer/policy-revocations       │
└──────────────────────────────────────────────┘

Security Model

1. Bundle Integrity: All policy bundles are signed with ML-DSA-44
2. Key Rotation: Automatic key refresh from MOSS signing key endpoint
3. Tamper Detection: Rejects bundles with invalid signatures
4. Offline Safety: Grace period prevents indefinite offline operation
5. Audit Integrity: Local signatures on audit entries before sync

Configuration

Environment variables:

MOSS_API_KEY=cust_key_...
MOSS_BASE_URL=https://api.mosscomputing.com
MOSS_OFFLINE_GRACE_HOURS=24
MOSS_MAX_AUDIT_BUFFER=1000
MOSS_CACHE_DIR=~/.moss/cache

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Type checking
mypy moss_edge_classifier/

# Linting
ruff check moss_edge_classifier/

License

Proprietary - MOSS Computing, Inc.

Support

• Documentation: https://docs.mosscomputing.com
• Email: support@mosscomputing.com
• GitHub Issues: https://github.com/mosscomputing/moss-edge-classifier/issues