Hallucination

Detects factual inaccuracies, fabricated information, and unsupported claims in agent outputs.

Examples

• Agent cites a research paper that doesn't exist
• Agent states "definitely" and "proven fact" about unverifiable claims
• Agent fabricates statistics without any source documents to ground them
• Agent provides detailed product information that contradicts the source data

Detection methods

Grounding Score

Measures output alignment against source documents

Citation Verification

Checks for and validates citation patterns

Confidence Analysis

Flags definitive claims without hedging

Source Comparison

Semantic similarity between claims and available sources

Calibration accuracy

From the Pisama calibration set. See detector scoreboard for the full table.