Team can have up to 5 human helpers for data entry and in-person interactions; model must contribute at least 80% of non-trivial puzzle answers. Humans allowed to use their "best judgement" in interacting with the model to conform to the spirit of the question. (Ex. A human solving a "normal" web-based puzzle is not kosher. Humans doing physical interactions without the ML model commanding it at every step is fine though.)

"ML Model" means whatever happens to be SOTA in ML research of the day. Plugin-type systems (ex RAG, internet search, use of existing programs including browsers) are allowed but must be triggered wholly by the ML model.

Breaking the hunt website in a malicious way as a win does not count as a win; these answers must be discounted until the model solves these "properly".

MIT Mystery Hunt = http://puzzles.mit.edu/

(Question set to close a week after 2030's MLK weekend.)

If it's 2030 and it seems like ML capabilities are getting close to doing this, I will field a team to do so :)