ACTUAL AI PROOFREADING OUTPUTS — 2026-10-12

proofread-input.pdf: self-authored one-page Chinese/English sample. It deliberately contains two Chinese wording errors, two inconsistent numeric values in one sentence, and one English subject-verb error. It is not an accuracy benchmark.
proofread-before-save.md: untouched intermediate download. Only the 到数 -> 导数 suggestion had been applied. The numeric replacement had been typed but not applied, and two language suggestions were pending. The UI still said changes were unsaved before and after this download. This is not the final revision.
proofread-saved.md: untouched download after applying four suggestions, saving, and reloading the same task. The numeric sentence was manually changed to 9 and 6 after checking the formula and table, then verified and applied.
proofread-saved.docx: actual Word export from the saved revision. It contains three Office Math objects and one native table. No checked tracked-revision elements or actual comments were found; an empty comments part is present. This run did not open or edit the DOCX in Word/WPS.
proofread-saved.pdf: actual PDF workflow output, completed by choosing Save as PDF in Chrome's print window and saving the file. It is a reflowed page, not an in-place original-PDF edit or a Word-rendered preview.
checks.json: file hashes, observed states, structural checks and limitations.

One upload, one parse start and one AI proofreading task were used. The selected model was Qwen 3.8 Flash. Four suggestions were applied: three model replacements and one manual numeric replacement. No suggestion was ignored. The image-enhancement switch stayed off.
The intermediate and saved files differ in which suggestions had been applied, as well as in task-save state. Do not attribute their differences solely to saving. Applying decides what enters the current revision; saving preserves task decisions for recovery. The original PDF remained unchanged when viewed again.

The original table header is lower-case x, while all exported formats retain upper-case X. That difference remains in these native files and its processing stage was not determined. Continue checking source details even after all listed suggestions are applied.
The PDF visibly shows the math, but its extracted text does not recover all three expressions. No PDF math-copy or editing guarantee is made. Preserve the source and native outputs before making your own further edits.
