OAIS Sentinel Technical Documentation

1. Quickstart (Get Running in 60 Seconds)

Install the sovereign Sentinel client directly via pip:

pip install oais-sentinel

Activate your tenant using your portable activation token obtained from oais.ai/register or AWS Marketplace:

sentinel init --token oais_act_eyJ0ZW5hbnRfaWQiOi...

Wrap your Amazon Bedrock runtime calls with a single line change:

from oais_sentinel import SentinelBedrockClient

# Drop-in replacement for boto3.client('bedrock-runtime')
bedrock = SentinelBedrockClient()

response = bedrock.invoke_model(
    modelId="anthropic.claude-3-5-sonnet-20241022-v2:0",
    body={"prompt": "Assess commercial portfolio sensitivity under OSFI E-23..."}
)

# Access cryptographic audit receipt
receipt = response["sentinel_receipt"]
print(f"Verdict: {receipt['gate_decision']}, Input Hash: {receipt['input_hash']}")

2. Zero-Knowledge Cryptographic Architecture

Unlike traditional logging gateways that collect and store raw customer queries, OAIS Sentinel enforces strict mathematical **Zero-Knowledge Data Custody**:

• Client-Side Salted Hashing: Prompts and completions are hashed locally in your application's memory using a client-held HMAC secret: SHA-256(content || client_salt).
• No Plain-Text Transmission: Raw prompts, financial records, and PII never touch OAIS servers or external networks.
• Append-Only Merkle Ledger: Only irreversible 64-character hexadecimal hashes, model versions, timestamps, and policy gate decisions are sequenced into the cryptographic tree.

3. Python Bedrock SDK (`SentinelBedrockClient`)

The SentinelBedrockClient provides complete feature parity with standard boto3 Bedrock runtime clients, with built-in enterprise resilience:

• Sub-Millisecond Overhead: Client-side cryptographic processing completes in 0.02ms – 0.48ms.
• Guaranteed Fail-Open Resilience: If local disk buffers or networks experience congestion, inference calls return immediately without disrupting production user sessions.
• AWS Partner PRM Attribution: Automatically decorates requests with AWS Partner Revenue Measurement tags (APN_1.1/...) for seamless AWS EDP commit drawdown.

4. Command-Line Interface Reference (`sentinel`)

The sentinel CLI provides complete offline operational and audit control:

`sentinel status`

Inspects local workspace configuration or cryptographic ledger state:

sentinel status

`sentinel verify`

Executes an independent first-principles audit of the cryptographic Merkle tree offline:

sentinel verify --db path/to/ledger.db

`sentinel cert`

Generates an auditor-ready compliance dossier (HTML & JSON) for OSFI examiners:

sentinel cert --db path/to/ledger.db

`sentinel cockpit`

Launches the Client-Side Tenant Governance Cockpit (Streamlit UI on port 8501) for real-time risk radar, RFC 6962 Merkle proof inspection, live prompt sandbox, and 1-click OSFI E-23 certificate export:

sentinel cockpit --port 8501

`sentinel ops`

Launches the OAIS Internal Fleet Operations Control Plane (port 8502) for multi-tenant usage metering, AWS Marketplace MMS batch export, regional Nitro Enclave health radar, and anomaly monitoring:

sentinel ops --port 8502

`sentinel proxy`

Launches a local zero-knowledge HTTP reverse proxy gateway for non-Python environments (Node.js, Go, Java):

sentinel proxy --port 8080

5. AWS Nitro Enclaves Hardware Attestation

For institutions requiring the highest level of regulatory defensibility, Sentinel anchors Merkle tree state heads (STHs) inside **AWS Nitro Enclaves**:

• Hardware Isolation: Nitro Enclaves have no persistent storage, no interactive access, and no external networking other than a secure local vsock channel.
• Cryptographic Attestation (PCR0): The enclave signs ledger roots using private keys sealed to the enclave's unique cryptographic measurement (PCR0 hash).
• Tamper Proof: Even AWS administrators or compromised host instances cannot forge an attestation signature.

6. Regulatory Compliance Mapping

How Sentinel satisfies specific regulatory requirements:

• OSFI Guideline E-23 (Model Risk Management): Provides deterministic proof of model lineage, tamper-evident risk intervention records, and independent auditor verification.
• SOC 2 Type II: Satisfies Security, Availability, and Confidentiality criteria through cryptographic proof custody.
• EU AI Act (Article 12): Fully satisfies high-risk AI system logging and traceability mandates.