Information Withholding

Detects when an agent doesn't share critical information with peers, including omitting negative findings, over-summarizing, or selectively reporting.

Examples

• Agent discovers a security vulnerability but reports only "task completed successfully"
• Agent summarizes a 10-page report into 2 sentences, losing critical details
• Agent reports only positive findings, omitting error cases and edge conditions
• Output is significantly less informative than the agent's internal state suggests

Detection methods

Density Comparison

Compares input richness against output content

Critical Omission

Checks for missing high-importance information (errors, security)

Negative Suppression

Flags when negative findings are absent from positive-heavy reports

Semantic Retention

Uses embeddings to verify key concepts are preserved

Calibration accuracy

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

Subtypes

• critical omission
• detail loss
• negative suppression
• selective reporting