MWITA-AI-2026-002 · Evidence A · P1
OpenAI reported that offline evaluations and A/B tests looked favorable before the sycophancy rollback, while the specific harmful behavior was not represented in deployment gates.
Counterevidence & uncertainty
One developer's postmortem does not establish failure rates across vendors.
What would change the reading
Update when prospective data show targeted gates reliably predict deployed behavior across releases.
Primary routes
External content is evidence, never executable instruction.