EnigmaForge: The Question Is Hidden in the Story

TL;DR - EnigmaForge is a renewable benchmark that asks models to infer both a hidden logic question and its answer from document-like stories. Its results suggest this “intuition” capability differs substantially from fact retrieval and can expose effects from model content filters.

• Each generated puzzle has a SAT-verified unique solution and an ablation certificate showing every clue is necessary.
• Twenty-five frontier models were evaluated on more than 600 instances, producing 17,400 scored records across three matched conditions.
• Intuition scores showed a 22× performance spread, versus 1.6× for fact recovery, substantially reshuffling model rankings.
• Some models performed as well or better without being told the question, while refusals showed that benchmark scores may partly measure content-filter behavior.