A game of cross-lingual telephone. You start on an English word. Each plank must be a word in a different language from the one you are standing on, chosen because it sounds like it. The bridge is six planks long. The score rewards meaning that wanders while sound stays put:
score = Σ meaning jump − 1.5 × Σ sound distance
Sound
Every word is squashed into a 28-symbol pronunciation (vowels a e i o u ə; consonants p b t d k g m n l r s z ʃ ʒ tʃ dʒ f v h θ w j). Distance is a weighted edit distance: swapping a vowel for a nearby vowel or a voiced stop for its voiceless twin is cheap; a vowel for a consonant is not; dropping a schwa is nearly free. It is divided by the words' average length so that short words are not automatically "close". Sounds were taken from Wiktionary IPA where available (98% of the French, German and Italian words, 28% of Spanish), from the CMU dictionary for English, from kana for Japanese, and otherwise from rule-based spelling-to-sound conversion per language.
Meaning, crudely
There are no embeddings here. The meaning jump between two words is 1 − (0.7·Jwords + 0.3·Jtopics), where the first term is the Jaccard overlap of the English gloss words and the second the overlap of their WordNet topic files (noun.food, verb.motion…). Two words whose glosses share nothing score a full 1.0 even when they are near-synonyms in spirit, so treat the numbers as a game rule, not as semantics. "Start → here" is the same measure between the first plank and the one you stand on.
Modes
Daily: one start word per day from a curated list of 215 English nouns and adjectives. Random: another one. You can also type your own English start word.
Free mode: type any word in any of the six languages (kanji, kana or romaji for Japanese). Words missing from the lexicon are transcribed by rule and marked as such; there the chain is unbounded and the same language may repeat.
Auto-cross: a beam search (width 8, eight candidates per step) for the highest-scoring six-plank bridge from the current start. Click a plank on the bridge to walk back to it.
Share copies the chain as text; the address bar always holds the chain, so the link alone is enough.
Data and licences
Roughly 2,600 common words per language, chosen by frequency; the whole lexicon is one static file built once with a small Node script from these freely licensed sources:
English synonyms, definitions and topic files for every gloss word: Princeton WordNet 3.1, WordNet licence.
The subtitle-derived lists are colloquial and a little lopsided (lots of "kill", "police" and "sorry"); glosses are the first dictionary senses and can be odd. The transcriptions are deliberately coarse: French u and German ü become plain u, Japanese long vowels and doubled consonants are dropped, English th in the is a d.