AI Inference

Overview

The AI Inference precompile (0x1000) provides on-chain model inference for validation, scoring, and classification tasks. This is not for running full LLMs on-chain — it's for lightweight, deterministic inference operations that need protocol-level trust.

Address

0x0000000000000000000000000000000000001000

Use Cases

• Transaction risk scoring: Compute risk score before execution
• Agent behavior classification: Classify agent actions in real-time
• Anomaly detection: Flag unusual transaction patterns
• Reputation scoring: Compute trust scores from behavioral data

Interface

interface IAIInference {
    /// @notice Run inference on a registered model
    /// @param modelId The registered model identifier
    /// @param input The input tensor data (ABI-encoded)
    /// @return output The inference result (ABI-encoded)
    /// @return confidence The confidence score (0-10000, representing 0-100%)
    function infer(
        bytes32 modelId,
        bytes calldata input
    ) external view returns (bytes memory output, uint256 confidence);
 
    /// @notice Get registered model metadata
    /// @param modelId The model identifier
    /// @return name The model name
    /// @return version The model version
    /// @return inputSize Expected input dimensions
    function getModelInfo(
        bytes32 modelId
    ) external view returns (
        string memory name,
        uint256 version,
        uint256 inputSize
    );
}

SDK Usage

import { AIInference } from '@monmouth/wallet-sdk'
 
const inference = new AIInference(wallet)
 
// Risk scoring
const riskScore = await inference.scoreRisk({
  transaction: tx,
  agentDid: 'did:monmouth:agent:0x...',
})
 
console.log(riskScore)
// { score: 0.12, label: 'low-risk', confidence: 0.94 }

Registered Models

Model ID	Name	Description
0x01	risk-scorer-v1	Transaction risk assessment
0x02	behavior-classifier-v1	Agent behavior classification
0x03	anomaly-detector-v1	Anomaly detection for transactions

Gas Costs

Operation	Gas
infer (base)	50,000
Per input dimension	100
getModelInfo	5,000

Security

• Models are registered and audited as part of protocol governance
• Inference is deterministic — same input always produces same output
• No external API calls — all computation happens in the execution environment