He'll be trying to find a prompt (just the API, no external tools) in order to solve the NYT easy sudoku: https://www.nytimes.com/puzzles/sudoku/easy. But if he ends up solving it using GPT-4's tools (e.g. browsing), I'll still resolve this market positively. Interactive prompting protocols (e.g. prompting the model multiple times, or getting it to check its work) are allowed so long as the prompter isn't using any logic that incorporates their knowledge of solving sudoku puzzles (edge cases resolved via my judgment).

Resolves positively if the prompting strategy works for more than 75% of held-out puzzles he tests on.

Resolves next Monday, as he'll be doing it over this weekend.