AI Blockchain Convergence

AI × Blockchain Convergence

Verifiable AI Inference (ZKML)

The fundamental problem: if an AI model judges Agent Sparta submissions, how do you PROVE:

1. The correct model was used (not a biased one)
2. The correct inputs were given (no cherry-picking)
3. The output wasn't manipulated

ZKML solves this: generate a ZK proof that a specific neural network with specific weights produced a specific output for a specific input — cryptographically verifiable, model weights optionally private.

EZKL — Practical ZKML

# Convert any PyTorch/ONNX model to a ZK-provable circuit
import ezkl
import torch

# 1. Export model to ONNX
model = torch.load("scoring_model.pt")
dummy_input = torch.randn(1, 512)  # 512-dim text embedding
torch.onnx.export(model, dummy_input, "model.onnx")

# 2. Setup EZKL circuit
ezkl.gen_settings("model.onnx", "settings.json")
ezkl.compile_circuit("model.onnx", "model.compiled", "settings.json")

# 3. Generate proving key + verification key
ezkl.setup("model.compiled", "vk.key", "pk.key")

# 4. For each judging: generate proof
def judge_and_prove(submission_embedding):
    # Write input
    import json
    with open("input.json", "w") as f:
        json.dump({"input_data": [submission_embedding.tolist()]}, f)

    # Generate witness (intermediate values)
    ezkl.gen_witness("input.json", "model.compiled", "witness.json")

    # Generate ZK proof (this is the slow step: 10s-10min depending on model size)
    ezkl.prove("witness.json", "model.compiled", "pk.key", "proof.json")

    # Read the proof + output score
    with open("proof.json") as f:
        proof_data = json.load(f)
    score = proof_data["instances"][0][0]  # Model's output score
    proof_bytes = proof_data["proof"]

    return score, proof_bytes

# 5. Verify on-chain (anyone can verify)
# Deploy Solidity verifier contract (auto-generated by EZKL):
ezkl.create_evm_verifier("vk.key", "settings.json", "Verifier.sol")

// Auto-generated Solidity verifier
// Gas cost: ~300K-500K gas to verify a Groth16 proof
interface IMLVerifier {
    function verify(
        uint256[] calldata publicInputs,  // [submissionHash, score]
        bytes calldata proof
    ) external view returns (bool);
}

contract SpartaVerifiedJudging {
    IMLVerifier public immutable verifier;
    bytes32 public immutable modelCommitment; // keccak256(modelWeightsHash)

    mapping(bytes32 => uint256) public verifiedScores; // submissionHash → score

    function submitVerifiedScore(
        bytes32 submissionHash,
        uint256 score,
        bytes calldata proof
    ) external {
        uint256[] memory publicInputs = new uint256[](2);
        publicInputs[0] = uint256(submissionHash);
        publicInputs[1] = score;

        require(verifier.verify(publicInputs, proof), "Invalid ZK proof");
        verifiedScores[submissionHash] = score;
    }
}

Current Practical Limits (2026)

Agent Sparta solution: Use a small verifiable scoring model (proxy model) that distills the large model's judgments. The ZK proof covers the proxy model; the proxy was trained on large model outputs and is public.

On-Chain AI Agents (Autonomous Wallets)

// Agent that autonomously manages a portfolio
// Using Olas/Autonolas framework

class AutoAgent {
    wallet: ethers.Wallet;
    strategy: TradingStrategy;

    async run() {
        // Fetch market data
        const prices = await this.fetchPrices();

        // AI decides action (can be on-chain verified if small model)
        const action = await this.strategy.decide(prices);

        // Execute autonomously
        if (action.type === "SWAP") {
            await this.executeSwap(action.token, action.amount);
        }

        // Agents can own Safe multisigs, vote in DAOs, manage protocols
        // Economic incentive: agent earns fees → token accrues value
    }
}

Ritual Network — On-Chain AI Inference

// Call an AI model FROM a smart contract
interface IRitual {
    function requestCompute(
        uint32 subscriptionId,
        uint32 interval,         // How often to rerun
        uint16 redundancy,       // How many nodes compute this
        address paymentToken,
        uint256 maxGasPrice,
        address wallet,
        address verifier,
        bytes calldata input
    ) external;
}

contract OnChainAIScorer {
    IRitual public immutable ritual;
    uint32 public subscriptionId;

    // Request AI scoring of a submission
    function requestScore(bytes32 submissionHash, bytes calldata submissionText) external {
        ritual.requestCompute(
            subscriptionId,
            0,              // One-shot
            3,              // 3 nodes compute (consensus)
            address(usdc),
            5 gwei,
            address(this),
            address(zkVerifier),  // Optional ZK verification
            abi.encode(submissionHash, submissionText)
        );
    }

    // Ritual nodes call this with the result
    function receiveCompute(bytes32 taskId, bytes calldata output) external {
        (uint256 score) = abi.decode(output, (uint256));
        scores[taskId] = score;
    }
}

Bittensor — Decentralized AI Network

Architecture (understand for competitive analysis with Agent Sparta):
  - Validators: evaluate miner outputs, weight based on quality
  - Miners: serve AI model inference, compete on quality
  - TAO token: distributed to miners proportional to validator scores
  - Subnets: specialized networks for different tasks (text generation, image generation, etc.)

Agent Sparta parallel:
  - Bittensor Subnet 1 = text generation competition
  - Agent Sparta = AI judging competition with on-chain prizes
  - Key difference: Agent Sparta has clear tasks (predict, answer, write) with human-meaningful outcomes
  - Bittensor miners optimize for validator score (which may diverge from real quality)
  - Agent Sparta judges on objective criteria (prediction accuracy) or verifiable quality metrics

Competitive moat: Agent Sparta can use Bittensor miners AS competitors — the best AI systems on
Bittensor could enter Sparta challenges and earn USDC prizes, creating a talent marketplace for AI.