Every board word and every legal clue is a 300-dimensional GloVe word vector. Pick the cards you want your partner to find; the browser ranks all 25 board words by cosine similarity to each of the ~7,000 legal clue words and reports the single clue that best separates your set.
Green means some legal word ranks all of your targets above every other card. Red means none does. That is the wall. The number is the projection margin: how far the best clue is from cleanly separating the set.
The second tab, Decoder-perfect, reader fails, replays real audits: clues the decoder rated perfect that a live language-model or human reader still couldn't recover. Decoder-recoverable and humanly readable come apart.
A 25-word board, a hidden subset, and one legal clue word. What can it communicate?
Explore the wall
Decoder-perfect, reader fails
Click words to pick a hidden target subset. The browser runs the
embedding-rank decoder over every legal clue and shows the best one.
Select one or more words.
blue = your targets ·
green = decoder picks it (correct) ·
red = decoder picks it instead (intruder) ·
amber = your target the decoder misses
blue = the set the decoder says this clue recovers ·
after reveal: green hit ·
red reader's wrong pick ·
amber intended word the reader missed