Nine languages found four bugs in English

Eloist now speaks nine languages. Seven are finished end to end, and by end to end we mean the whole product: English, Spanish, German, French, Chinese, Japanese and Russian. Hebrew and Arabic have the full interface and are still finishing their written content, so they are offered exactly as far as they are ready and no further.

Nobody picks that list. Each language earns its release from its own content, counted: the interface, the emails, the opening guides, the stories, the chess lore. A half-finished language cannot ship by accident, and a finished one cannot sit invisible. The second half is the one we care about, because it had already happened to us. French was complete, every dictionary, and it was offered nowhere, because back then a language was only on when somebody remembered to add it to a list.

Russian went from nothing to fully released in 93 minutes, timed from the first command to the check against the live site. That covered 5,349 interface keys and 1,189 email keys as the dictionaries stood that day, the opening guides and the stories and the lore, and 3,174 opening names. Where a player's surname had no settled Russian spelling we left it in Latin letters rather than guess at it. The coach films are narrated in your language now too: six that had spoken English at Spanish and French members were re-rendered in the right one.

Translating a product is a test of it

Here is the part we did not expect. Four of the things we found had nothing to do with translation. They were ordinary bugs, they had been going out to English readers for weeks, and the only reason anyone saw them is that the language changed. Every place the code assumed English, it was really assuming something else and getting away with it.

The sentence that asks you to reply

Seven of the fourteen most recent films were missing their last line from the audio. In English too. Always the same line, the one that asks you to write back. The speech model sometimes stops early and still reports success, and the check meant to catch that forgave exactly one missing word at the end. That allowance was calibrated on English, where the closing phrase is two words, so dropping one still tripped it. Spanish puts the imperative and both pronouns into a single word, “dímelo”, and losing it costs the audio one token and the film its entire ending.

Then we counted, across those fourteen films: the gap was zero seven times and two or more seven times. Not once was it exactly one. The tolerance was protecting nothing that happens and hiding something that does. It is zero now.

The coach spoke French. The screen did not.

Every on-screen card in every film was hardcoded English, sitting in a Python file that not one of our translation checks had ever opened. A French member heard the coach in French and read “Was this useful?” on the screen. The labels on the board were English on every frame. Two smaller faults came out of the same file. A layout tuned to the width of an English word, which is why a French board label ran off the edge of the frame. And a test that looked for the English word “lost” inside a sentence the writer translates, so a visual cue quietly vanished from every film that was not in English.

The envelope, and the greeting

The third fault was the wrapper around every email. Whatever language the letter inside was written in, the frame around it was English, including the one letter a paying member is guaranteed to receive. Hebrew and Arabic letters were laid out left to right, which is the wrong way round for both.

The fourth is the one we like least. The coach opened Spanish films with an English greeting, because a prompt told the model, in English, to keep a literal English phrase. A model was told to say an English thing, so it said it. No scanner can catch that, because it is not a template anywhere. It is an instruction.

What we cannot tell you yet

No native speaker has read any of the nine. Every mechanical check is green, and every one of those checks is about shape: are the keys there, do the dates format, does the page run right to left, is the last word of the script in the audio. None of them can tell you whether a sentence sounds like a person wrote it. If you read Eloist in a language that is not English and something lands badly, that is the report we want most.