Proofreading and offline use: frequently asked questions

The proofreading code processes text in your browser and does not send drafts to cloud AI, the hosting platform, or model download services. Network requests still fetch the app, checking runtime, and models you choose to prepare. Hosting services may process connection and performance data. See the privacy and offline notice for the distinction.

Do not rely on recovery after a close or refresh. Drafts, imported originals, and undo history stay in page memory and are not saved automatically. Copy or export a TXT file before leaving or choosing Update and reload. The app cache is not a draft backup.

No account is needed. The app saves your interface language preference, not your drafts, IP address, or country information. It may also cache app and prepared model files. For details, including language settings retained from older versions, see what is stored and cached.

No. It covers common typos, repetition, punctuation, and a small set of bounded grammar rules. MacBERT mainly replaces individual Chinese characters; optional Qwen can attempt broader corrections and light polishing but may make incorrect edits. See real Chinese rule examples and their boundaries.

You can enter it, but the current Chinese rules and default model mainly target Simplified Chinese. Equal correction quality is not guaranteed, and drafts are not automatically converted to Simplified Chinese. Review Traditional Chinese names and wording yourself; no suggestions does not prove the text is correct.

Quick check uses local Harper with American English by default. There is currently no dialect selector. It needs no MacBERT or Qwen download, although your first visit must download the app and checking runtime. See an English correction example and the checking limits.

Yes. Quick check combines Chinese rules with English Harper checks. The interface language controls buttons and descriptions only: changing it does not translate your draft, change the English dialect, or download a model.

No. Complex grammar, context, facts, names, and specialist terms still need human review. Optional style suggestions are preferences rather than definite errors. If the app reports a failed check, resolve the failure or retry; a failure is not a clean result.

Quick check accepts up to 20,000 characters per draft. Local AI allows up to 3,000 characters and 80 sentences per check. Paste plain text or import and export TXT; direct Word/PDF import and their formatting are not supported. Split longer text into sections. See length and file limits.

You can accept, reject, or edit each suggestion. Finalize shows a read-only preview before you confirm actionable pending edits; rejected suggestions stay rejected. Ambiguous notices without a safe replacement need manual review. Follow the full-document review steps.

Undo restores the change and its review state, and you can inspect the imported original. Manual edits invalidate affected old suggestions, so use the results of a new check. Undo history belongs to the current page and cannot recover a draft from a closed page.

Choose Ignore spelling in this draft on an English spelling suggestion; you can restore it in Check settings. Only the same complete word with the same capitalization is ignored, and other grammar checks still run. This list is not saved when you close the page.

No files are available in a browser that has never prepared them. Open the app online and confirm the app and quick-check files are ready in Privacy & offline. Models need separate preparation at your request; a page that opens offline does not prove AI is ready. Follow the offline setup steps.

Any removed app or model files need to be prepared again. Caches belong to the current browser and site, do not automatically sync to other devices, and may be evicted by the browser. Switching models does not automatically remove the other model's cache. You can inspect and clear individual caches in the app.

Quick check needs JavaScript and WebAssembly, not WebGPU. Local AI requires usable WebGPU, enough storage, and GPU memory. If those are unavailable, quick check remains available; the app does not switch to cloud AI. Use the app's device check and preparation results to confirm readiness.

The layout adapts to narrow screens, while a computer offers more space for long drafts. Models and offline use depend on the browser, device, and available memory. Real-phone acceptance testing has not been completed, so the same experience cannot be guaranteed on every phone.