Yay I'm glad the absurd humor resonated! That kind of goofiness makes me giggle, and the giggles are what keep me building :D. I'm not sure how much you want to know about the word association stuff, but here's an attempt to answer! Sorry for the wall of text.
I think you're asking about how the game knows which answers are partially correct (checkmark vs cross)? That is largely based on the Google NGram dataset . The 2-gram data lists the frequency of certain 2-word pairs [as found in books], which can be used to figure out which word pairs have relationships or special meaning, relative to other words. Around 2020, I wrote code to slurp that up and make it searchable via my Dillfrog Context Search website (example) . So for the jam, I used my internal API to search that data and write code that "bakes" level files from my manually-generated list of correct answers and clues.
To generate the puzzles (i.e. the "inputs" to the baking process), I manually read through a list of "ambiguous" words that had multiple meanings , in the hopes that they led to more interesting contexts and puzzles. For each word, I searched Wiktionary (primarily) and my Dillfrog Context data (secondarily) to see what words might be recognizable clues. Then I created a TSV file that included the intended correct answers, and the clues to use.
Then, to build the JSON "output" that the game will actually read and use, I run that baking process. The baker slurps the TSV. For each clue, it uses that Google NGram data to spit out the top ~500 words that occur before or after each clue (based on a PMI score if I remember correctly...), so it can acknowledge the partially-correct guesses [for the top ~500 results]. It also adds correct answers that I missed (e.g. plurals, if I listed the answer in singular form), based on the top ~500 connections, and does some other cleanup.
Hope that helps explain it!