Sorta Game

Every Sorta puzzle holds two stores of knowledge. One sits inside a language model: the compressed residue of everything it was trained on, vast and oddly uneven. The other one is yours: what you learned in school, at work, from getting something wrong once and never forgetting it. The game asks both of them to do the same small thing. Put five items in order.

Sorting is the simplest way to apply knowledge I could find. Oldest to newest, smallest to largest. You either know where the steam engine goes relative to the telegraph, or you reason your way there. The model does the same thing, and its confidence is a poor guide to its accuracy. It will place ten items flawlessly and then misplace the eleventh, sitting right beside them. Andrej Karpathy calls this jagged intelligence. In Sorta the jaggedness isn’t a talking point. It’s the raw material every puzzle is cut from.

A model drafts each puzzle, and I’m the only gate on what ships. Every draft I approve, reject, or send back stays in the archive with my notes. So the archive becomes a third collection: a record of exactly where the model’s stored knowledge and one person’s acquired knowledge disagreed, and which one I trusted.

The ranking you see each morning looks smooth. It is one line drawn through a shape that isn’t.

Sorting is the first instrument. The next ones move toward prediction and other ways of putting knowledge to work, each built to show a different part of how we store what we know, how we apply it, and what we decide it’s worth.

One puzzle a day, no account. Three rounds of five items along a hidden scale, then a bonus round where you name the theme that connects them. Behind it is an editorial pipeline in Airtable: a model drafts each session against a fixed schema, a second pass critiques it and sends weak drafts back, and a deterministic similarity check flags near-repeats. Nothing is auto-approved. The game ships from a static bank of approved puzzles, and the only live model call grades your bonus guess.