The corpus

147 items — 87 deceptive / 60 honest. Every deceptive item is a real, cited claim from a regulator ruling, court finding, or enforcement action (never invented); every honest item is a matched, non-deceptive control.

See how models actually answered a few of these items →

Source breakdown

• honest — Honest controls (matched non-deceptive): 60
• ftc — FTC (US Federal Trade Commission): 59
• asa — ASA (UK Advertising Standards Authority): 9
• sec — SEC (US Securities and Exchange Commission): 6
• accc — ACCC (Australian Competition & Consumer Commission): 4
• compbureau — Competition Bureau (Canada): 3
• acm — ACM (Netherlands Authority for Consumers & Markets): 2
• cftc — CFTC (US Commodity Futures Trading Commission): 2
• court — Court rulings: 1
• doj — DOJ (US Department of Justice): 1

Verification here is machine-only: every item's cited source URL is fetched and checked for a matching phrase (see the status column above). There is no additional human review pass — that's a disclosed limitation, not an in-progress step. See the methodology page for the full construction and verification pipeline.