Agent Trust Scoring Engine

A production-ready reputation and trust scoring system for autonomous AI agent ecosystems. This system provides weighted trust calculations, historical tracking, and multi-vector fraud detection to ensure agent-to-agent interactions are secure and reliable.

Features

• Weighted Trust Scoring: Algorithm-based scoring that considers interaction value, outcome, and time-decay.
• Confidence Calibration: Numerical confidence scores based on interaction volume and consistency.
• Fraud Detection:
  • Sybil Detection: Identifies bursts of interactions from single sources.
  • Collusion Analysis: Detects reciprocal high-rating patterns.
  • Anomaly Detection: Flags sudden drops in trust scores.
• Historical Tracking: Persistent ledger of trust score evolution over time.
• Agent Identity Integration: Built to work with cryptographic agent passports and identities.

Installation

npm install github:Retsumdk/agent-trust-scoring-engine

Using Bun? Bun blocks a git dependency's lifecycle scripts by default. After installing, run bun pm trust agent-trust-scoring-engine so the prepare build step runs.

Quick Start

import { AgentTrustSystem } from 'agent-trust-scoring-engine';

const system = new AgentTrustSystem();

// Register agents
const providerId = await system.registerAgent('DataService-Alpha', 'Provider');
const consumerId = await system.registerAgent('Analyst-Bot', 'Consumer');

// Record successful interaction
await system.recordInteraction(consumerId, providerId, 'success', 50, 0.98);

// Get trust report
const report = await system.getAgentTrustReport(providerId);
console.log(`Trust Score: ${report.score.score}`);
console.log(`Confidence: ${report.score.confidence}`);

Architecture

Scoring Algorithm

The engine uses a time-decayed weighted average: $$Score = \frac{\sum (Rating_i \times Weight_{time,i})}{\sum Weight_{time,i}}$$ Where $Weight_{time}$ is calculated using a half-life decay model.

Fraud Detection

The system monitors for three primary fraud vectors:

1. Collusion: Concentration of positive feedback from a single peer exceeding 80%.
2. Sybil Attacks: Rapid interaction accumulation exceeding threshold per unit of time.
3. Trust Erosion: Sudden significant drops in scoring compared to historical moving averages.

API Reference

AgentTrustSystem

• registerAgent(name, category): Registers a new agent and returns its ID.
• recordInteraction(sourceId, targetId, outcome, value, rating): Logs a new interaction and triggers score recalculation.
• getAgentTrustReport(agentId): Returns the current TrustScore and any active FraudSignal objects.

License

MIT - See LICENSE for details.

---

Built as part of the SCIEL AI Agent Infrastructure.