blindference-node 0.3.3

Blindference Node – Confidential inference worker for agentic AI.

Blindference Node

Confidential inference worker for the Blindference decentralized AI execution network.

Run encrypted inference jobs, earn fees, and help build a private, verifiable, and economically accountable AI execution layer on Arbitrum Sepolia.

---

What It Does

Blindference Node is the runtime that executes confidential inference tasks assigned by the Inference Coordination Layer (ICL). Each node:

• Attests its identity and capabilities to the ICL (mock TEE for tier 0, TPM/TEE for higher tiers)
• Heartbeats every 60 seconds to prove liveness (ICL) and every 10 days on-chain
• Polls for pending job assignments every 5 seconds
• Decrypts encrypted prompts via CoFHE threshold FHE under strict ACL
• Executes inference via Groq Llama 3, Google Gemini, or local vLLM (pluggable backends)
• Commits results back to the ICL for quorum consensus
• Earns fees for successful task completion

---

Quick Start

1. Install

pip install blindference-node

For GPU-accelerated local inference (optional):

pip install "blindference-node[gpu]"

Or install from source:

git clone https://github.com/AbhishekPanwarr/Blindference-node.git
cd Blindference-node
pip install -e .

2. Initialize

blindference-node init

This will:

1. Detect GPU capabilities (or default to mock inference)
2. Auto-detect cloud API keys (Groq / Gemini) from environment
3. Generate an encrypted Ethereum wallet keystore
4. Save configuration to ~/.blindference/config.json

Non-interactive mode:

export BLF_PRIVATE_KEY=0x...
export BLF_KEY_PASSWORD=secure_password
blindference-node init

3. Attest

# Interactive attestation (choose mock or TEE)
blindference-node attest

# Or skip interactive menu and use mock directly
blindference-node attest --mock

# With custom development key
blindference-node attest --mock --tee-key mydevkey

After ICL attestation, optionally register on-chain with gas estimation.

4. Stake BLIND Tokens

Nodes must stake at least 1000 BLIND to participate in inference quorums. Staking provides economic security — stake is slashed if a node produces incorrect output or times out.

# Stake 1000 BLIND (minimum)
blindference-node staking stake 1000

# Check your stake status
blindference-node staking status

# Initiate unstake (starts 96h unbonding)
blindference-node staking unstake

# Complete unstake after unbonding period
blindference-node staking withdraw

Staking Economics:

• Minimum stake: 1000 BLIND
• Unbonding period: 96 hours
• Reward per job: 1 BLIND (60% leader, 20% each verifier)
• Slashing: 3 consecutive failures → entire stake hard-slashed on-chain

5. Run

blindference-node run

The daemon starts four concurrent loops:

Loop	Frequency	Purpose
ICL Heartbeat	Every 60s	Proves liveness to ICL (free REST call)
On-Chain Heartbeat	Every 10 days	Proves liveness to NodeRegistry (gas tx)
Attestation Watchdog	Every 10min	Auto-re-attests if certificate expires within 6h
Assignment Poller	Every 5s	Polls ICL for pending inference jobs

---

Commands

Command	Description	Status
init	Initialize node — wallet, GPU detection, save config	✅ Ready
attest	Attest node with ICL (mock / TEE) and optionally register on-chain	✅ Ready
run	Start daemon — heartbeat, watchdog, job polling & execution	✅ Ready
status	Show node status — address, tier, models, cert expiry	✅ Ready
staking stake	Stake BLIND tokens to join inference quorums	✅ Ready
staking unstake	Initiate unstake (starts 96h unbonding)	✅ Ready
staking withdraw	Complete withdrawal after unbonding period	✅ Ready
staking status	Show BLIND stake status and failure count	✅ Ready
models list	List all registered inference backends and availability	✅ Ready
models test	Test a specific backend with a prompt	✅ Ready
models add	Register a custom backend from a dotted Python path	✅ Ready
test-determinism	Run GPU determinism self-test with vLLM or cloud APIs	✅ Ready
jobs list	List completed jobs and BLIND earnings	✅ Ready
jobs earnings	Total BLIND earned across all completed jobs	✅ Ready
balance	Current BLIND token balance	✅ Ready

---

Architecture

flowchart TB
    subgraph "Blindference Node"
        DAEMON[Daemon Process]
        HB[ICL Heartbeat<br/>60s]
        ONCHAIN[On-Chain Heartbeat<br/>10 days]
        WD[Attestation Watchdog<br/>10min]
        POLL[Assignment Poller<br/>5s]
        WORKER[Job Worker]
        
        DAEMON --> HB
        DAEMON --> ONCHAIN
        DAEMON --> WD
        DAEMON --> POLL
        POLL -->|dispatch| WORKER
    end
    
    subgraph "External Services"
        ICL[ICL Coordinator]
        COFHE[CoFHE Network]
        IPFS[IPFS Gateway]
        LLM[Groq / Gemini / vLLM]
        CHAIN[Arbitrum Sepolia]
    end
    
    HB -->|POST /internal/heartbeat| ICL
    ONCHAIN -->|NodeRegistry.heartbeat()| CHAIN
    WD -->|GET /internal/challenge| ICL
    WD -->|POST /internal/attestation/verify| ICL
    POLL -->|GET /internal/assignments/{addr}| ICL
    WORKER -->|POST /internal/task/claim| ICL
    WORKER -->|decryptForView| COFHE
    WORKER -->|download blob| IPFS
    WORKER -->|run inference| LLM
    WORKER -->|POST /internal/task/result| ICL
    
    ICL -->|register attestation| CHAIN

---

Configuration

All configuration is stored in ~/.blindference/config.json and can be overridden via environment variables prefixed with BLF_.

Default Config

{
  "node_address": "0x...",
  "keystore_path": "~/.blindference/keystore.json",
  "tier": 0,
  "supported_model_ids": ["facebook/opt-125m"],
  "custom_backends": [],
  "attestation_backend": "mock",
  "icl_endpoint": "https://icl.blindference.xyz",
  "fhenix_rpc": "https://testnet.fhenix.zone",
  "ipfs_gateway": "https://node.lighthouse.storage",
  "model_cache_dir": "~/.blindference/models",
  "log_level": "INFO",
  "network": "fhenix_testnet",
  "attestation_cert_hash": "",
  "attestation_expiry": 0,
  "registered_on_chain": false,
  "stake_amount_wei": 0,
  "cofhe_mode": "bridge",
  "cofhe_endpoint": "",
  "cofhe_chain_id": 421614,
  "skip_output_key_storage": false
}

Environment Variables

Variable	Type	Description
BLF_NODE_ADDRESS	string	Ethereum address (0x...)
BLF_KEYSTORE_PATH	string	Path to encrypted keystore
BLF_TIER	int	Attestation tier (0=mock, 1=TPM, 2=TEE)
BLF_SUPPORTED_MODEL_IDS	list	Comma-separated model IDs
BLF_CUSTOM_BACKENDS	list	Dotted Python paths for custom backends
BLF_ATTESTATION_BACKEND	string	mock, tpm, or sgx
BLF_ICL_ENDPOINT	string	ICL base URL
BLF_FHENIX_RPC	string	EVM RPC endpoint
BLF_IPFS_GATEWAY	string	IPFS download/upload gateway
BLF_LOG_LEVEL	string	DEBUG, INFO, WARNING, ERROR
BLF_COFHE_MODE	string	bridge (TypeScript subprocess) or python (HTTP)
BLF_COFHE_ENDPOINT	string	CoFHE/EVM RPC endpoint
BLF_COFHE_CHAIN_ID	int	Chain ID for CoFHE (421614 for Arbitrum Sepolia)
BLF_KEY_PASSWORD	string	Keystore decryption password
GROQ_API_KEY	string	Enables Groq cloud backend
GOOGLE_API_KEY	string	Enables Gemini cloud backend
MOCK_ATTESTATION_KEY	string	Override default mock attestation key

CoFHE Modes

bridge (default): Spawns a TypeScript subprocess via @cofhe/sdk/node for CoFHE operations. More reliable, handles SDK lifecycle correctly.

python (alternative): Direct HTTP calls to CoFHE endpoints. Lighter weight but requires manual session management.

---

Model Backends

Blindference Node uses a pluggable backend registry to support multiple inference providers.

Built-in Backends

Backend	Description	Requirements	Model IDs
vllm	Local GPU inference	NVIDIA GPU + vllm package	facebook/opt-125m
groq	Groq cloud API	GROQ_API_KEY env var	groq:llama-3.3-70b-versatile
gemini	Google Gemini REST API	GOOGLE_API_KEY env var	gemini:gemini-2.5-flash
mock	Deterministic SHA-256 fallback	Always available	* (universal fallback)

Custom Backends

from blindference_node.models.base import ModelBackend

class MyBackend(ModelBackend):
    def name(self) -> str: return "my-backend"
    def is_available(self) -> bool: return True
    def supported_models(self) -> list[str]: return ["my-model"]
    def run(self, model_id: str, prompt: str) -> str: return "result"

Register via CLI:

blindference-node models add my_package.backends:MyBackend

---

Node Lifecycle

1. Initialization (init)

sequenceDiagram
    participant User
    participant Node as blindference-node
    
    User->>Node: init
    Node->>Node: detect GPU / cloud API keys
    Node->>Node: generate wallet + keystore
    Node->>Node: save config

2. Attestation (attest)

sequenceDiagram
    participant Node as blindference-node
    participant ICL
    participant Chain as Arbitrum Sepolia
    
    Node->>ICL: GET /internal/challenge/{addr}
    ICL-->>Node: challengeId + nonce
    Node->>Node: generate attestation quote
    Node->>ICL: POST /internal/attestation/verify
    ICL-->>Node: certHash + expiry + tier
    opt On-chain registration
        Node->>Chain: NodeRegistry.register()
        Chain-->>Node: tx hash
    end
    Node->>Node: save config

3. Daemon Execution (run)

sequenceDiagram
    participant Node as blindference-node
    participant ICL
    participant Chain as Arbitrum Sepolia
    participant COFHE as CoFHE Network
    participant LLM as Groq/Gemini
    
    Node->>Node: check attestation expiry
    alt expired or missing
        Node->>ICL: auto-re-attest
    end
    
    par ICL Heartbeat
        loop Every 60s
            Node->>ICL: POST /internal/heartbeat
        end
    and On-Chain Heartbeat
        loop Every 10 days
            Node->>Chain: NodeRegistry.heartbeat()
        end
    and Watchdog
        loop Every 10min
            Node->>Node: check cert expiry < 6h
            alt needs re-attest
                Node->>ICL: re-attest
            end
        end
    and Assignment Poller
        loop Every 5s
            Node->>ICL: GET /internal/assignments/{addr}
            alt assignments found
                Node->>Node: spawn job worker (max 2 concurrent)
            end
        end
    end

4. Job Execution

sequenceDiagram
    participant Worker as Job Worker
    participant ICL
    participant COFHE as CoFHE Network
    participant IPFS
    participant LLM as Groq/Gemini
    
    Worker->>ICL: POST /internal/task/claim
    ICL-->>Worker: kpHighHandle + kpLowHandle + claimDeadline
    
    alt leader role
        Worker->>COFHE: decryptForView(kpHighHandle).withPermit()
        Worker->>COFHE: decryptForView(kpLowHandle).withPermit()
        Worker->>Worker: reconstruct AES key
        Worker->>IPFS: download encrypted prompt blob
        Worker->>Worker: AES decrypt prompt
        Worker->>LLM: run inference
        Worker->>Worker: AES encrypt output
        Worker->>Worker: split output key into halves
        Worker->>COFHE: encrypt output key halves
        Worker->>Chain: storeKey(output key) in PromptKeyStore
        Worker->>IPFS: upload encrypted output blob
        Worker->>ICL: POST /internal/task/result
    else verifier role
        Worker->>COFHE: decryptForView(kpHighHandle).withPermit()
        Worker->>COFHE: decryptForView(kpLowHandle).withPermit()
        Worker->>Worker: reconstruct AES key
        Worker->>IPFS: download encrypted prompt blob
        Worker->>Worker: AES decrypt prompt
        Worker->>LLM: run inference
        Worker->>Worker: hash result
        Worker->>ICL: POST /internal/task/verify
    end

---

Requirements

• Python: 3.10 or higher
• Operating System: Linux (tested on Ubuntu 22.04), macOS
• GPU (optional): NVIDIA GPU with CUDA support for local vLLM inference
• Network: Outbound HTTPS to ICL, Fhenix RPC, IPFS gateway, Groq/Gemini APIs
• Storage: ~2GB for model cache (if running local models)

---

Development

Setup

# Clone repository
git clone https://github.com/AbhishekPanwarr/Blindference-node.git
cd Blindference-node

# Create virtual environment
python -m venv venv
source venv/bin/activate

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

# Run tests
pytest tests/ -v

Project Structure

Blindference-node/
├── blindference_node/          # Main package
│   ├── __init__.py
│   ├── cli.py                  # CLI entry points (init, attest, run, status, models)
│   ├── node_loop.py            # Daemon with heartbeat, watchdog, poller
│   ├── job_handler.py          # Task execution logic
│   ├── crypto.py               # CoFHE client, AES blob encryption/decryption
│   ├── icl_client.py           # ICL REST API client
│   ├── wallet.py               # Ethereum wallet generation and loading
│   ├── registry.py             # On-chain registration and heartbeat
│   ├── config.py               # Configuration management
│   ├── attestation/            # Attestation backends (mock, TPM, TEE)
│   ├── models/                 # Pluggable inference backends
│   │   ├── base.py
│   │   ├── vllm_backend.py
│   │   ├── groq_backend.py
│   │   ├── gemini_backend.py
│   │   ├── mock_backend.py
│   │   └── registry.py
│   ├── backend_loader.py       # Dynamic backend loading
│   └── scripts/                # TypeScript CoFHE bridge
│       └── cofhe_bridge.mjs
├── tests/                      # Test suite
│   ├── test_crypto.py
│   ├── test_job_handler.py
│   ├── test_node_loop.py
│   ├── test_icl_client.py
│   ├── test_registry.py
│   ├── test_e2e.py
│   └── test_wallet.py
├── contracts/                  # Solidity contract ABIs
├── pyproject.toml              # Package configuration
├── docker-compose.yml          # Docker orchestration
├── Dockerfile                  # Container image
└── README.md                   # This file

Docker

# Build image
docker build -t blindference-node .

# Run container
docker run -d \
  --name blindference-node \
  -e BLF_KEY_PASSWORD=secure_password \
  -e BLF_ICL_ENDPOINT=https://icl.blindference.xyz \
  -v /host/config:/root/.blindference \
  blindference-node run

---

Security Model

Tier 0 (Mock Attestation)

• Software-only attestation with no hardware trust
• Quote: HMAC-SHA256 of nonce + runtime code hash
• For development only — nodes can be impersonated

Tier 1 (TPM 2.0)

• TPM-backed attestation with measured boot
• Hardware-bound identity
• Requires TPM 2.0 chip

Tier 2 (TEE / SGX / TDX)

• Intel SGX or AMD SEV enclave attestation
• Confidential computing — code and data encrypted in memory
• Remote attestation verified by ICL against manufacturer quoting enclaves

Slashing Conditions

Nodes may be slashed for:

• Failed attestation: Missing or expired certificate
• Missed heartbeat: No ICL heartbeat within 5 minutes
• Bad inference: Verifier consensus shows wrong result
• Timeout: Failed to submit result within execution window

---

Documentation

• Quickstart Guide — Step-by-step first node setup
• Attestation Guide — Mock, TPM, and TEE attestation
• Configuration — All config options and env vars
• Model Backends — Pluggable backend system
• Rewards & Slashing — How nodes earn and what gets them slashed
• Troubleshooting — Common issues and fixes

---

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

---

License

MIT License — see LICENSE for details.

---

Consensus Calibration Feedback

Every accepted inference job feeds into a long-term calibration loop that improves quorum accuracy over time:

Signal	Source	Impact
Thumbs Up	User decrypts output, votes "accurate"	Increases confidence weight for leader's model provider
Thumbs Down	User decrypts output, votes "not accurate"	Flags potential false positives; triggers threshold review
Dispute Override	User overrides quorum rejection	Signals verifier pool may be too strict
Verdict Mismatch	Verifiers disagree with accepted leader	Reduces reputation of outlier verifier

Nodes benefit from accurate calibration — a well-tuned quorum means fewer false rejections (more jobs accepted, more fees earned) and fewer false acceptions (less slashing risk).

Support

• Docs: docs.blindference.xyz
• Discord: Blindference Community
• Twitter: @blindference
• Issues: GitHub Issues