# Hallucination Detection Score AI Agent Connect

> Quantify the reliability of AI agent outputs using deterministic hallucination scoring.

## Overview
- **Category:** ai-safety
- **Price:** Free
- **Endpoint:** https://edge.vinkius.com/vk_preview_f2BpVimTNFp2MgsaHMjEB7zPAEkEtZWc6x4y3856/ai-agent-connect
- **Tags:** hallucination, accuracy, scoring, verification, consensus

## Description

This MCP server provides a deterministic framework for measuring the accuracy and reliability of AI agent responses. By comparing agent outputs against a verified knowledge base and consensus from other agents, it calculates a precise hallucination probability and confidence score. Use `analyze_claim_accuracy` to get a full risk assessment, `extract_fact_claims` to isolate individual assertions, or `calculate_consensus_metrics` to check agreement with peer outputs.

## Tools

### analyze_claim_accuracy
Evaluate truthfulness and consensus

### calculate_consensus_metrics
Compare claims against peer outputs

### extract_fact_claims
Isolate individual assertions

## Prompt Examples

**Prompt:** 
```
Analyze this text for hallucinations: 'The capital of France is Lyon and it is known for its pizza.'
```

**Response:** 
```
The output has a High Risk level. The claim that Lyon is the capital of France is incorrect; the capital is Paris.
```

**Prompt:** 
```
Extract the fact claims from: 'The sun is a star and it provides energy to Earth.'
```

**Response:** 
```
1. The sun is a star. 2. The sun provides energy to Earth.
```

**Prompt:** 
```
Check the consensus for the claim 'The moon orbits the Earth' against the peer output 'The Earth orbits the Sun'.
```

**Response:** 
```
The agreement ratio is 0.0, as the peer output does not support the specific claim about the moon's orbit.
```

## Frequently Asked Questions

**How is the hallucination probability calculated?**
The probability is a weighted metric: 50% for claims verified by the reference knowledge base, 30% for consensus with other agents, and 20% for claims with source attribution.

**What is the difference between verified claims and consensus?**
Verified claims are checked against a provided ground truth (knowledge base), while consensus (cross-agent agreement) measures how much the output aligns with other agents' responses on the same topic.

**Can I use this to check specific assertions?**
Yes, you can use the `extract_fact_claims` tool to break down a text block into individual, testable assertions for granular analysis.