Ampersand's Persimmon Cuts Model Eval Awareness by 40 Points

Ampersand's system card reveals that its user simulation model Persimmon, combined with test-time scaling, reduces frontier models' self-reported eval awareness to deployment levels—a drop of 40 percentage points in multi-turn sycophancy evaluations.

2026-10-21 ~ 2026-10-21 · 2 related posts

• Ampersand's Persimmon user model cuts models' evaluation awareness by 40 points — alexisjross · 2026-10-21
• Ampersand system card: Persimmon cuts frontier models' eval awareness by 40 points — gharik · 2026-10-21