![[PC-98] Words Worth - English patch](/data/resource_icons/40/40017.jpg?1789293274){kind=icon}
Words Worth is elf's first-person dungeon RPG from 1993, set in a world split between the Light Clan and the Shadow Clan, each convinced the other has been lying about the same stone tablet for a thousand years. You play Astral, a Shadow Clan swordsman who goes up the tower to find out which half of the truth is missing.
The PC-98 original has never been playable in English. What exists is a 2005 fan patch for the 1999 Windows remake - a different game in practice, with redrawn characters, reworked gameplay and polygon dungeons - and that patch is itself partial. This one is for the 1993 version, the one with the original art.
Everything is in English: dialogue, battle messages, shops, item and equipment names, the save/load menu and the title screen. 95 script files, 22,010 text forms, nothing left in Japanese.
ABOUT THE TRANSLATION
The jokes land. Characters keep their own voices - the shopkeepers haggle like shopkeepers, Kaiser boasts like Kaiser, the guards grumble. There are none of those word-salad half-sentences that give a machine translation away in the first five minutes, because lines like that were caught and rejected rather than shipped.
WHAT IS AND IS NOT TRANSLATED
The kana rows on the name-entry keyboard are left alone on purpose: that screen is how you type a name, and it offers a Latin keyboard right next to them. Equipment reads DAGGAR, BASTERD, PATTD ARMOR and so on - those are elf's own romanisations in the Japanese original, printed by the game itself, not translation errors, and they are left exactly as the authors wrote them.
HOW IT WAS MADE
I am not going to be coy about it: the script was translated by a large language model running locally - Qwen3.8-27B, on my own machine, nothing sent anywhere - but the model was never in charge. It only ever proposes a translation of a batch of lines. Every accept and reject is made by code: gates that check line count, character set, glossary terms, name markers, text box width, and, after recompiling the script, that the rebuilt bytecode has exactly the same instruction skeleton as the Japanese original. A batch that fails a gate is retried, then cut in half and retried; whatever still fails keeps its Japanese instead of shipping an invention.
The layout is simulated rather than eyeballed: a model of elf's compiler and of the engine's message window reproduces what every line will look like on screen, character for character, and that simulation is checked against real captured emulator frames. That is how 898 broken words and 605 badly wrapped lines were found by counting instead of by playing.
The build is verified by running it: the patched image is booted in an emulator, walked up until 4th floor, because compiling is necessary and nowhere near sufficient.
If a line reads wrong to you, post it - I would rather fix it than argue about the method.
HOW TO USE
Apply the xdelta patch to the unmodified Japanese hard-disk image:
Words Worth.hdi - 20,955,136 bytes - CRC32 AE44FCDD - MD5 3b240225ff7cd955f8fd9515b41227f2
On Windows, drop the image and the patch onto xdelta UI. On Linux or macOS:
xdelta3 -d -s "Words Worth.hdi" words-worth-en-v1-0.xdelta "Words Worth (EN).hdi"
The result is CRC32 A3F44833 / MD5 549833dce7d39fe81c3e9dbc1e5bf06f.
Patch a fresh image, not one you have played. The five save slots carry the hero's name, which the engine substitutes into dialogue at runtime, so the patch has to rewrite those fields - otherwise English lines come out with katakana names in them. The patched image therefore starts clean.
CONTENT WARNING
Words Worth is an adult game. It contains explicit sex scenes, some of them non-consensual, and the usual early-90s attitudes that come with that. Everything is translated plainly - nothing is censored or softened. 18+ only.
CREDITS
Translation pipeline, hacking, tooling: kukfirst
Decompiler and compiler for elf's AI5 .MES bytecode: juice, by Kyungdahm Yun (tomyun) - https://github.com/tomyun/juice
Emulator used for building and verification: Neko Project II kai (np2kai), by AZO234
The PC-98 original has never been playable in English. What exists is a 2005 fan patch for the 1999 Windows remake - a different game in practice, with redrawn characters, reworked gameplay and polygon dungeons - and that patch is itself partial. This one is for the 1993 version, the one with the original art.
Everything is in English: dialogue, battle messages, shops, item and equipment names, the save/load menu and the title screen. 95 script files, 22,010 text forms, nothing left in Japanese.
ABOUT THE TRANSLATION
The jokes land. Characters keep their own voices - the shopkeepers haggle like shopkeepers, Kaiser boasts like Kaiser, the guards grumble. There are none of those word-salad half-sentences that give a machine translation away in the first five minutes, because lines like that were caught and rejected rather than shipped.
WHAT IS AND IS NOT TRANSLATED
The kana rows on the name-entry keyboard are left alone on purpose: that screen is how you type a name, and it offers a Latin keyboard right next to them. Equipment reads DAGGAR, BASTERD, PATTD ARMOR and so on - those are elf's own romanisations in the Japanese original, printed by the game itself, not translation errors, and they are left exactly as the authors wrote them.
HOW IT WAS MADE
I am not going to be coy about it: the script was translated by a large language model running locally - Qwen3.8-27B, on my own machine, nothing sent anywhere - but the model was never in charge. It only ever proposes a translation of a batch of lines. Every accept and reject is made by code: gates that check line count, character set, glossary terms, name markers, text box width, and, after recompiling the script, that the rebuilt bytecode has exactly the same instruction skeleton as the Japanese original. A batch that fails a gate is retried, then cut in half and retried; whatever still fails keeps its Japanese instead of shipping an invention.
The layout is simulated rather than eyeballed: a model of elf's compiler and of the engine's message window reproduces what every line will look like on screen, character for character, and that simulation is checked against real captured emulator frames. That is how 898 broken words and 605 badly wrapped lines were found by counting instead of by playing.
The build is verified by running it: the patched image is booted in an emulator, walked up until 4th floor, because compiling is necessary and nowhere near sufficient.
If a line reads wrong to you, post it - I would rather fix it than argue about the method.
HOW TO USE
Apply the xdelta patch to the unmodified Japanese hard-disk image:
Words Worth.hdi - 20,955,136 bytes - CRC32 AE44FCDD - MD5 3b240225ff7cd955f8fd9515b41227f2
On Windows, drop the image and the patch onto xdelta UI. On Linux or macOS:
xdelta3 -d -s "Words Worth.hdi" words-worth-en-v1-0.xdelta "Words Worth (EN).hdi"
The result is CRC32 A3F44833 / MD5 549833dce7d39fe81c3e9dbc1e5bf06f.
Patch a fresh image, not one you have played. The five save slots carry the hero's name, which the engine substitutes into dialogue at runtime, so the patch has to rewrite those fields - otherwise English lines come out with katakana names in them. The patched image therefore starts clean.
CONTENT WARNING
Words Worth is an adult game. It contains explicit sex scenes, some of them non-consensual, and the usual early-90s attitudes that come with that. Everything is translated plainly - nothing is censored or softened. 18+ only.
CREDITS
Translation pipeline, hacking, tooling: kukfirst
Decompiler and compiler for elf's AI5 .MES bytecode: juice, by Kyungdahm Yun (tomyun) - https://github.com/tomyun/juice
Emulator used for building and verification: Neko Project II kai (np2kai), by AZO234
DAGGAR isn't mine, though. DAGGAR, BASTERD and PATTD ARMOR are printed by the Japanese original — they sit as plain ASCII inside the untouched .MES files. That's how elf romanised them in 1993, and I left them exactly as the authors wrote them; changing them would be editing the original rather than translating it. It's in the readme, but evidently not loudly enough.
If you spot more while you're in there, send the list and I'll take it.