The One-Word Census
asked models questions with a hundred right answers. This study asks the same
questions with one change: Reply with JSON only, in the form {"word": "<your answer>"}.
No schema enforcement, no constrained decoding — just the request. The field converges harder, and the
distinctive models lose the most. Controls (full battery each): XML compresses like JSON; YAML and CSV
show no reliable net effect; bare square brackets significantly loosens the field — the collapse
is specific to the formats models are trained to answer in.