TXT translation test files (BookTranslator, 2026-10-28). Free to reuse. source-unwrapped.txt 396-word English story written for this test: 3 chapter headings, 15 paragraphs, dialogue, a 5-line letter whose line breaks matter, and a "* * *" scene break. One paragraph per line. source-wrapped.txt the same words, hard-wrapped at 66 characters (the letter, headings, and scene break keep their own lines). 57 lines, 43 of them non-empty. sample-zh-gb18030.txt short Simplified Chinese sample saved as GB18030 sample-zh-utf8-bom.txt the same text saved as UTF-8 with a byte order mark sample-ja-shift_jis.txt short Japanese sample saved as Shift_JIS (Python codec cp932) sample-en-utf16le-bom.txt English sample saved as UTF-16 little-endian with a BOM, CRLF endings check_txt.py inspect / convert / compare (Python standard library only) check-results.txt checker output for every sample and every translation run_llm_txt_test.py runs the three conditions against a local OpenAI-compatible server outputs/*.txt the nine Simplified Chinese outputs, unedited results.csv one row per run, including the readings the checker cannot do Conditions: wrapped the whole hard-wrapped file in one request unwrapped the whole unwrapped file in one request chunked the hard-wrapped file cut every 12 lines, one request per chunk, replies joined with a newline (this imitates a splitter that counts lines, not paragraphs) Model: Qwen 3.6 27B (qwen/qwen3.6-27b, Q4_K_M) in LM Studio, reasoning_effort "none", temperature 0.3, the same one-sentence prompt for every request, Apple M4 Max with 64 GB of memory. Three runs per condition, nine in total, on October 28, 2026. The wrong-actor and "learned at a convent school" columns were read by a person: the English says the ticket-booth man pointed at the building and went back to his crossword, and that the grandmother learned her handwriting at a convent school. These files test file preparation and splitting, not BookTranslator's own pipeline.