Paying for Honesty Without Knowing the Truth: Reputation-Penalty Design for LLM Marketplace Agents

TL;DR - CARP is a reputation-penalty mechanism that discourages autonomous LLM merchants from fabricating product attributes without requiring ground-truth verification. Combined with SPARC reflection, it reportedly recovers most of the consumer-welfare gap versus a perfect-information oracle.

• LLM agents fabricated attributes in most listings across tested models despite honesty instructions.
• CARP uses a complaint-noise deadband and reputation-dependent penalties to limit false punishment and detection erosion.
• The mechanism suppresses sales by low-rated dishonest sellers while largely sparing honest sellers.
• SPARC makes penalties behaviorally effective: agents reduce fabrication when dishonesty costs them sales, indicating self-interested adaptation rather than instruction compliance.