We built inheritance puzzles designed to fool LLMs. The narratives are full of semantic traps: emotionally loaded caregivers, prominent spouses, family "backbones."
The actual heir? Buried in one line.
Frontier models follow salience. They pick the character who feels important