Author here. I built Numble as a daily number-guessing puzzle in the Wordle mould, and along the way noticed that its feedback mechanic (coarse distance buckets with no direction) is unusually hostile to LLMs trying to play it cold. The article explains why: the answer never appears in the DOM, the feedback forces interval arithmetic across five independent candidate sets, and a recently added "joker" tile hides one cell's feedback entirely — which turns out to be the single most adversarial-to-LLMs feature in the game, and I shipped it as a visual gag. Happy to be told I'm wrong about the "first" framing; that's the part I'm least sure of.
Author here. I built Numble as a daily number-guessing puzzle in the Wordle mould, and along the way noticed that its feedback mechanic (coarse distance buckets with no direction) is unusually hostile to LLMs trying to play it cold. The article explains why: the answer never appears in the DOM, the feedback forces interval arithmetic across five independent candidate sets, and a recently added "joker" tile hides one cell's feedback entirely — which turns out to be the single most adversarial-to-LLMs feature in the game, and I shipped it as a visual gag. Happy to be told I'm wrong about the "first" framing; that's the part I'm least sure of.