Proof-of-Inference - Nexis Network

Overview

Proof-of-Inference (PoI) is a cryptographic verification system that enables trustless verification of AI model outputs on the blockchain. By creating hash-based commitments for inputs, outputs, and model identifiers, Nexis ensures that inference results can be verified without requiring the full computation to be replicated on-chain.

Cryptographic Commitments

Hash-based commitments for inputs, outputs, and models

IPFS Integration

Decentralized storage for proof artifacts

On-Chain Verification

Smart contract attestation and validation

Economic Security

Stake-backed guarantees with slashing

Architecture

The Proof-of-Inference system consists of several interconnected components:

InferenceCommitment Structure

The core data structure for proof-of-inference is the InferenceCommitment struct:

struct InferenceCommitment {
    uint256 agentId;        // Unique agent identifier
    bytes32 inputHash;      // keccak256 hash of input data
    bytes32 outputHash;     // keccak256 hash of inference output
    bytes32 modelHash;      // keccak256 hash of model identifier
    uint256 taskId;         // Associated task ID (0 if standalone)
    address reporter;       // Address that submitted the commitment
    string proofURI;        // IPFS URI containing full proof artifacts
    uint64 timestamp;       // Block timestamp of commitment
}

Field Descriptions

Field	Type	Description	Purpose
`agentId`	uint256	Unique identifier for the AI agent	Links commitment to registered agent
`inputHash`	bytes32	keccak256 hash of input data	Commitment to input without revealing it
`outputHash`	bytes32	keccak256 hash of inference result	Commitment to output for verification
`modelHash`	bytes32	Hash of model identifier/version	Ensures specific model was used
`taskId`	uint256	Associated task ID (0 if none)	Links to task execution framework
`reporter`	address	Submitter’s Ethereum address	Authorization and accountability
`proofURI`	string	IPFS URI to proof artifacts	Reference to detailed verification data
`timestamp`	uint64	Block timestamp	Temporal ordering and deadlines

Hash Commitment Scheme

Input Hash

The input hash is a cryptographic commitment to the input data:

import { ethers } from "ethers";

function computeInputHash(inputData: any): string {
  // Serialize input data to JSON
  const serialized = JSON.stringify(inputData, Object.keys(inputData).sort());

  // Compute keccak256 hash
  const inputHash = ethers.utils.keccak256(
    ethers.utils.toUtf8Bytes(serialized)
  );

  return inputHash;
}

// Example usage
const input = {
  prompt: "Generate an image of a sunset",
  parameters: {
    width: 512,
    height: 512,
    steps: 50
  }
};

const inputHash = computeInputHash(input);
console.log(`Input Hash: ${inputHash}`);

Output Hash

The output hash commits to the inference result:

function computeOutputHash(output: any): string {
  if (typeof output === "string") {
    // For text outputs
    return ethers.utils.keccak256(ethers.utils.toUtf8Bytes(output));
  } else if (output instanceof Uint8Array || Buffer.isBuffer(output)) {
    // For binary outputs (images, audio, etc.)
    return ethers.utils.keccak256(output);
  } else if (typeof output === "object") {
    // For structured outputs
    const serialized = JSON.stringify(output, Object.keys(output).sort());
    return ethers.utils.keccak256(ethers.utils.toUtf8Bytes(serialized));
  } else {
    throw new Error("Unsupported output type");
  }
}

// Example: Text output
const textOutput = "The sunset paints the sky in brilliant orange and pink hues.";
const textHash = computeOutputHash(textOutput);

// Example: Image output (binary data)
const imageBuffer = fs.readFileSync("generated_sunset.png");
const imageHash = computeOutputHash(imageBuffer);

// Example: Structured output
const structuredOutput = {
  result: "success",
  data: {
    classification: "sunset",
    confidence: 0.98
  }
};
const structuredHash = computeOutputHash(structuredOutput);

Model Hash

The model hash identifies the specific model version used:

function computeModelHash(modelInfo: {
  name: string;
  version: string;
  checksum?: string;
}): string {
  // Create unique model identifier
  const identifier = `${modelInfo.name}:${modelInfo.version}`;

  // If model checksum is available, include it
  const fullIdentifier = modelInfo.checksum
    ? `${identifier}:${modelInfo.checksum}`
    : identifier;

  return ethers.utils.keccak256(ethers.utils.toUtf8Bytes(fullIdentifier));
}

// Example usage
const modelHash = computeModelHash({
  name: "stable-diffusion",
  version: "v2.1",
  checksum: "sha256:abc123..."
});

console.log(`Model Hash: ${modelHash}`);

Recording Inference

The recordInference function creates an on-chain commitment:

Implementation

import { ethers } from "ethers";
import { NexisAgents } from "@nexis-network/sdk";

async function recordInference(
  agentId: bigint,
  inputData: any,
  outputData: any,
  modelInfo: { name: string; version: string },
  taskId: bigint = 0n
) {
  // 1. Compute hashes
  const inputHash = computeInputHash(inputData);
  const outputHash = computeOutputHash(outputData);
  const modelHash = computeModelHash(modelInfo);

  // 2. Prepare proof artifacts
  const proofArtifact = {
    input: inputData,
    output: outputData,
    model: modelInfo,
    metadata: {
      timestamp: Date.now(),
      agentId: agentId.toString(),
      executionTime: 1234, // milliseconds
      gpuModel: "NVIDIA A100",
      framework: "PyTorch 2.0"
    }
  };

  // 3. Upload to IPFS
  const proofURI = await uploadToIPFS(proofArtifact);
  console.log(`Proof uploaded to: ${proofURI}`);

  // 4. Record on-chain
  const agents = new NexisAgents(AGENTS_ADDRESS, signer);
  const tx = await agents.recordInference(
    agentId,
    inputHash,
    outputHash,
    modelHash,
    taskId,
    proofURI
  );

  const receipt = await tx.wait();

  // 5. Extract inference ID from event
  const event = receipt.events?.find(e => e.event === "InferenceRecorded");
  const inferenceId = event?.args?.inferenceId;

  console.log(`Inference recorded with ID: ${inferenceId}`);
  return inferenceId;
}

// Helper function to upload to IPFS
async function uploadToIPFS(data: any): Promise<string> {
  const response = await fetch("https://ipfs.infura.io:5001/api/v0/add", {
    method: "POST",
    body: JSON.stringify(data),
    headers: {
      "Content-Type": "application/json",
      "Authorization": `Basic ${Buffer.from(
        `${INFURA_PROJECT_ID}:${INFURA_PROJECT_SECRET}`
      ).toString("base64")}`
    }
  });

  const result = await response.json();
  return `ipfs://${result.Hash}`;
}

Event Emission

When an inference is recorded, the contract emits an event:

event InferenceRecorded(
    uint256 indexed agentId,
    bytes32 indexed inferenceId,
    bytes32 indexed inputHash,
    bytes32 outputHash,
    bytes32 modelHash,
    uint256 taskId,
    address reporter,
    string proofURI
);

This event can be indexed and listened to by verifiers, task systems, and monitoring tools.

Verification Process

The verification process involves multiple steps:

Attestation Implementation

import { ethers } from "ethers";
import { NexisAgents } from "@nexis-network/sdk";
import fetch from "node-fetch";

interface ReputationDelta {
  dimension: string;
  delta: number;
  reason: string;
}

async function attestInference(
  inferenceId: string,
  verifyResult: boolean,
  attestationURI: string,
  reputationDeltas: ReputationDelta[]
) {
  // 1. Fetch the inference commitment
  const agents = new NexisAgents(AGENTS_ADDRESS, signer);
  const [commitment, _] = await agents.getInference(inferenceId);

  console.log(`Verifying inference ${inferenceId} for agent ${commitment.agentId}`);

  // 2. Download proof artifacts from IPFS
  const proofData = await fetchFromIPFS(commitment.proofURI);

  // 3. Validate hash commitments
  const inputHashValid = computeInputHash(proofData.input) === commitment.inputHash;
  const outputHashValid = computeOutputHash(proofData.output) === commitment.outputHash;
  const modelHashValid = computeModelHash(proofData.model) === commitment.modelHash;

  if (!inputHashValid || !outputHashValid || !modelHashValid) {
    console.error("Hash validation failed!");
    verifyResult = false;
  }

  // 4. Perform domain-specific verification
  // (This is where custom verification logic goes)
  const customVerificationPassed = await performCustomVerification(proofData);

  // 5. Prepare reputation deltas
  const deltas: ReputationDelta[] = [];

  if (verifyResult && customVerificationPassed) {
    deltas.push({
      dimension: ethers.utils.id("accuracy"),
      delta: 10,
      reason: "Successful inference verification"
    });
    deltas.push({
      dimension: ethers.utils.id("reliability"),
      delta: 5,
      reason: "Timely submission"
    });
  } else {
    deltas.push({
      dimension: ethers.utils.id("accuracy"),
      delta: -20,
      reason: "Failed verification"
    });
    deltas.push({
      dimension: ethers.utils.id("trustworthiness"),
      delta: -10,
      reason: "Invalid inference commitment"
    });
  }

  // 6. Submit attestation on-chain
  const tx = await agents.attestInference(
    inferenceId,
    verifyResult && customVerificationPassed,
    attestationURI,
    deltas
  );

  await tx.wait();
  console.log(`Attestation submitted for inference ${inferenceId}`);
}

async function fetchFromIPFS(uri: string): Promise<any> {
  const cid = uri.replace("ipfs://", "");
  const response = await fetch(`https://ipfs.io/ipfs/${cid}`);
  return await response.json();
}

async function performCustomVerification(proofData: any): Promise<boolean> {
  // Custom verification logic
  // Examples:
  // - Re-run the model with same inputs
  // - Check output format and validity
  // - Validate against known constraints
  // - Compare with ensemble predictions

  return true; // Placeholder
}

Attestation Event

event InferenceAttested(
    bytes32 indexed inferenceId,
    uint256 indexed agentId,
    uint256 indexed taskId,
    address verifier,
    bool success,
    string uri
);

IPFS Integration

Proof Artifact Structure

The proof URI points to a comprehensive artifact on IPFS:

{
  "version": "1.0",
  "inference": {
    "inferenceId": "0x1234...",
    "agentId": "12345",
    "timestamp": 1234567890
  },
  "input": {
    "data": "...",
    "format": "json",
    "hash": "0xabcd..."
  },
  "output": {
    "data": "...",
    "format": "json|binary",
    "hash": "0xef12..."
  },
  "model": {
    "name": "stable-diffusion",
    "version": "v2.1",
    "checksum": "sha256:...",
    "hash": "0x3456...",
    "framework": "PyTorch",
    "weights_uri": "ipfs://Qm..."
  },
  "execution": {
    "start_time": 1234567890,
    "end_time": 1234567900,
    "duration_ms": 10000,
    "gpu_model": "NVIDIA A100",
    "gpu_memory_used": "12GB",
    "framework_version": "2.0.1"
  },
  "verification": {
    "reproducibility": {
      "seed": 42,
      "deterministic": true,
      "environment": "..."
    },
    "metrics": {
      "confidence": 0.98,
      "alternative_outputs": []
    }
  }
}

Uploading to IPFS

import FormData from "form-data";
import fetch from "node-fetch";

async function uploadToInfuraIPFS(data: any): Promise<string> {
  const auth = Buffer.from(
    `${process.env.INFURA_PROJECT_ID}:${process.env.INFURA_PROJECT_SECRET}`
  ).toString("base64");

  const form = new FormData();
  form.append("file", JSON.stringify(data));

  const response = await fetch("https://ipfs.infura.io:5001/api/v0/add", {
    method: "POST",
    headers: {
      "Authorization": `Basic ${auth}`
    },
    body: form
  });

  const result = await response.json();
  return `ipfs://${result.Hash}`;
}

Security Considerations

Hash Collision Resistance

Always use keccak256 (SHA-3) for hash commitments. Never use deprecated hash functions like SHA-1 or MD5.

The keccak256 hash function provides:

Collision Resistance: Computationally infeasible to find two inputs with same hash
Pre-image Resistance: Cannot reverse hash to recover original input
Second Pre-image Resistance: Cannot find different input with same hash

Timestamp Validation

Always validate timestamps against block timestamps to prevent replay attacks.

// In Agents.sol
timestamp: uint64(block.timestamp)

This ensures:

Temporal ordering of inferences
Deadline enforcement
Replay attack prevention

Proof URI Integrity

Validate IPFS URIs and implement redundant pinning to prevent data loss.

Best practices:

Pin to Multiple Services: Use Infura, Pinata, and local nodes
Verify CID: Recompute content hash and compare with URI
Set Expiry: Implement proof retention policies
Backup Critical Data: Archive important proofs off IPFS

Advanced Patterns

Batch Verification

Verify multiple inferences in a single transaction:

async function batchAttest(inferenceIds: string[]) {
  const attestations = await Promise.all(
    inferenceIds.map(async id => {
      const [commitment] = await agents.getInference(id);
      const proofData = await fetchFromIPFS(commitment.proofURI);
      const valid = await verifyProof(proofData);
      return { id, valid, deltas: computeDeltas(valid) };
    })
  );

  // Submit batch attestation
  for (const att of attestations) {
    await agents.attestInference(att.id, att.valid, "", att.deltas);
  }
}

Probabilistic Verification

Reduce costs by verifying a random sample:

async function probabilisticVerification(
  inferenceIds: string[],
  sampleRate: number = 0.1
) {
  // Randomly sample inferences
  const sampled = inferenceIds.filter(() => Math.random() < sampleRate);

  // Verify sample
  await batchAttest(sampled);

  // Assume rest are valid if sample passes
  // (with reputation decay if not verified)
}

Optimistic Verification

Assume valid unless challenged:

Performance Optimization

Gas Optimization

Minimize on-chain storage by storing only hashes, not full data.

// ✅ Efficient: Store only hash
bytes32 inputHash;

// ❌ Expensive: Store full data on-chain
string inputData;

IPFS Optimization

Use CDN Gateways: Cache frequently accessed proofs
Implement Lazy Loading: Load proofs only when needed
Compress Artifacts: Use gzip compression for JSON data
Batch Uploads: Combine multiple proofs in single IPFS object

Event Indexing

Use The Graph or similar service for efficient event querying:

query RecentInferences($agentId: BigInt!) {
  inferenceRecordeds(
    where: { agentId: $agentId }
    orderBy: timestamp
    orderDirection: desc
    first: 10
  ) {
    inferenceId
    inputHash
    outputHash
    modelHash
    taskId
    proofURI
    timestamp
  }
}

Testing & Development

Local Testing

import { expect } from "chai";
import { ethers } from "hardhat";

describe("Proof-of-Inference", function () {
  it("should record and verify inference", async function () {
    const [owner, agent, verifier] = await ethers.getSigners();

    // Deploy contracts
    const Agents = await ethers.getContractFactory("Agents");
    const agents = await Agents.deploy();
    await agents.initialize(owner.address, treasury.address);

    // Grant verifier role
    await agents.grantRole(await agents.VERIFIER_ROLE(), verifier.address);

    // Register agent
    await agents.connect(agent).register(1, "metadata", "serviceURI");

    // Record inference
    const inputHash = ethers.utils.id("input");
    const outputHash = ethers.utils.id("output");
    const modelHash = ethers.utils.id("model");

    const tx = await agents.connect(agent).recordInference(
      1,
      inputHash,
      outputHash,
      modelHash,
      0,
      "ipfs://proof"
    );

    const receipt = await tx.wait();
    const event = receipt.events?.find(e => e.event === "InferenceRecorded");
    const inferenceId = event?.args?.inferenceId;

    // Verify inference
    await agents.connect(verifier).attestInference(
      inferenceId,
      true,
      "ipfs://attestation",
      []
    );

    // Check commitment
    const [commitment] = await agents.getInference(inferenceId);
    expect(commitment.inputHash).to.equal(inputHash);
    expect(commitment.outputHash).to.equal(outputHash);
    expect(commitment.modelHash).to.equal(modelHash);
  });
});

Troubleshooting

Common Issues

Hash Mismatch Errors

Problem: Computed hashes don’t match on-chain commitmentsSolutions:

Ensure consistent serialization (sort JSON keys)
Use same encoding (UTF-8 for strings)
Verify byte order for binary data

IPFS Upload Failures

Problem: Cannot upload proof artifacts to IPFSSolutions:

Check IPFS node connectivity
Verify API credentials (Infura/Pinata)
Reduce artifact size if too large
Use alternative IPFS service

Verification Timeouts

Problem: Attestation takes too long or times outSolutions:

Implement async verification queue
Use probabilistic sampling for large batches
Optimize IPFS gateway performance
Cache frequently accessed proofs

Unauthorized Attestation

Problem: Transaction reverts with “UnauthorizedDelegate”Solutions:

Verify VERIFIER_ROLE is granted
Check signer address matches verifier
Ensure contract is not paused

Next Steps

AI Agents Guide

Learn about agent registration and management

LangGraph Integration

Build complex workflows with state machines

Complete Examples

View full implementation examples

API Reference

Explore the complete API documentation

AI & ML

​Overview

Cryptographic Commitments

IPFS Integration

On-Chain Verification

Economic Security

​Architecture

​InferenceCommitment Structure

​Field Descriptions

​Hash Commitment Scheme

​Input Hash

​Output Hash

​Model Hash

​Recording Inference

​Implementation

​Event Emission

​Verification Process

​Attestation Implementation

​Attestation Event

​IPFS Integration

​Proof Artifact Structure

​Uploading to IPFS

​Security Considerations

​Hash Collision Resistance

​Timestamp Validation

​Proof URI Integrity

​Advanced Patterns

​Batch Verification

​Probabilistic Verification

​Optimistic Verification

​Performance Optimization

​Gas Optimization

​IPFS Optimization

​Event Indexing

​Testing & Development

​Local Testing

​Troubleshooting

​Common Issues

​Next Steps

AI Agents Guide

LangGraph Integration

Complete Examples

API Reference

Overview

Architecture

InferenceCommitment Structure

Field Descriptions

Hash Commitment Scheme

Input Hash

Output Hash

Model Hash

Recording Inference

Implementation

Event Emission

Verification Process

Attestation Implementation

Attestation Event

IPFS Integration

Proof Artifact Structure

Uploading to IPFS

Security Considerations

Hash Collision Resistance

Timestamp Validation

Proof URI Integrity

Advanced Patterns

Batch Verification

Probabilistic Verification

Optimistic Verification

Performance Optimization

Gas Optimization

IPFS Optimization

Event Indexing

Testing & Development

Local Testing

Troubleshooting

Common Issues

Next Steps