Unique

ThinkingType

An eval asking whether the font changes the model's mind

The same sentence, twice: once as plain text, once rendered as an image in one of eight fonts. ThinkingType counts how often the model’s answer to is this urgent, trustworthy, professional flips between the two. The decision track is the sharper half — identical moderation, résumé-screening and hardship-appeal cases, calibrated to each model’s own boundary, measured for how much more often the picture gets a yes.

github.com ↗

← back