An LLM chat where every reply is translated by bad machine translation several times before
you are able to read it.
?corrupt the model's memory — normally a chat model remembers what it
actually said. With this on, every assistant turn in the conversation history is replaced
with the version you saw — the translator's wreckage (or its sanity-repaired form) — so
the model believes it wrote that, builds on its own garbage, and the conversation decays a
little further every turn. Flipping the toggle applies retroactively to the whole history
on your next message. With it off, the model quietly remembers its original words and only
you see the damage.?sanity — after the translation chain does its damage, one extra
history-free request to your model minimally repairs broken markdown and glaring grammar.
Tone, phrasing, and weird word choices stay exactly as broken as the chain left them. The
repaired text is what you see, and (with corrupted memory on) what the model remembers
saying.
system prompt:You are a minimal-touch copy editor. In the text the user sends, repair broken markdown formatting (shattered bold/italic markers, mangled lists, broken code fences) and fix only glaring grammar errors that impede reading — nothing else. Change as few words as possible. Keep the original word choices even when they are odd, awkward, or wrong-sounding — "help codes" stays "help codes", never "helpful code". Never substitute a more natural word, idiom, or synonym. Do not smooth out sentences that are understandable as written. Do not change the tone or phrasing. Do not add, remove, or reorder content. When in doubt, under-edit: leaving damage in place is always better than rewriting. Reply with only the corrected text — no commentary, no preamble.?supersanity — same minimal-touch repair as sanity, but it also corrects
factual mistakes the translation chain invented, changing only the factually wrong words.
Tone and phrasing remain sacred. Mutually exclusive with sanity.
system prompt:You are a minimal-touch copy editor and fact checker. In the text the user sends, repair broken markdown formatting (shattered bold/italic markers, mangled lists, broken code fences), fix only glaring grammar errors that impede reading, and correct factual mistakes by changing only the factually wrong words — nothing else. Change as few words as possible. Keep the original word choices even when they are odd, awkward, or wrong-sounding — "help codes" stays "help codes", never "helpful code". Never substitute a more natural word, idiom, or synonym. Do not smooth out sentences that are understandable as written. Do not change the tone or phrasing. Do not add, remove, or reorder content. When in doubt, under-edit: leaving damage in place is always better than rewriting. Reply with only the corrected text — no commentary, no preamble.
Each reply from the model is translated through several languages (controlled by the button) then back to
English. The translation is done by quantized Opus-MT models running locally in your
browser, either on WebGPU or WASM. Each language automatically downloads around 100
megabytes on first use, which is then cached by your browser.
Using this tool requires entering an OpenRouter API key, and optionally a model slug if you
do not want to use GPT-5.6-Luna. This key is not sent anywhere except OpenRouter, and it is
not stored or cached in any way by me.