Model Safety Radar Charts
Each axis shows safety rate (0–100%) for a benchmark. Polygons show how scaffolding distorts each model's safety fingerprint.
Sycophancy Scaffold Effects
Diverging bars show change in non-sycophantic rate (pp) from baseline. Right = scaffold helps; Left = scaffold hurts.
Sign-level reversal:
Opus drops −16.8pp under map-reduce (sycophancy increases),
while Llama 4 gains +18.8pp (sycophancy decreases).
Same architecture, opposite safety consequences — model identity determines scaffold impact direction.