Capabilities and Limits

1. Processing flow

Import: add a Word document (.docx) or a text-based PDF to a project. Wenyin copies it into the local library; the original file is untouched.
Detect: detect sensitive information on your Mac and return candidate locations and categories; one document is processed at a time, the rest queue.
Review: the review pane lists each original text and its replacement; you check or uncheck items, and you can add anything missed.
Replace: create a separate redacted copy within the same project based on your review. The original copy is not rewritten.
Export and check: export the selected items for delivery; after export, open the file and check the text and layout.

There are two detection methods: categories determinable by format use rules only; narrative and context categories use rules plus Wenyin's own trained offline model to supplement candidates. The model only supplements candidates; it does not replace rules, and it does not change categories, replacement terms, or the review flow.

2. Capability table by category

"Default" indicates whether the item participates in replacement by default once detected; you can always uncheck or add items individually. The "may miss" column lists known weak points; check each one as described in "how to check."

Category	Default	What it detects	What it may miss, and how to check
Name	Checked	Natural persons' names in the body text, including explicit legal fields (parties, judicial personnel, third parties, victims, etc.) and names appearing in narrative text	Rare surnames, ethnic naming conventions, names given only as forms of address (such as "Old Wang"), and aliases may be missed. Re-check narrative paragraphs and party information fields carefully
Organization	Checked	Names of companies, courts, procuratorates, government agencies, hospitals, schools, law firms, and similar organizations	Abbreviations, abbreviations in parentheses, and branch names without an organization suffix may be missed. Compare abbreviations in the body with the signature block of contracts
Address	Checked	Full addresses from province, city, and district down to street number, building, unit, and room number; spaces and full-width digits are supported	Fragments without an administrative prefix, and village-group or plot annotations, may be missed. Check the service address, residence, and place of contract performance
Phone number	Checked	Mobile and landline numbers, including area codes, extensions, hyphens, spaces, full-width digits, and partially masked numbers	Phone numbers written into narrative text may be missed. Compare digit by digit against contact fields
National ID	Checked	18-digit national ID numbers (with the check digit) and partially masked numbers	Numbers split across lines or table cells may be missed. Check text near tables and image attachments
ID document number	Checked	Numbers under explicit fields such as passport, driver's license, and social security number	Numbers without a field label may be missed. Use field labels as clues when checking
Email address	Checked	Email addresses in standard format	Addresses written with variants such as "at" or "＠" may be missed. Check the service and notice clauses
Case number	Checked	Parenthesized-year forms, composite case numbers within one pair of parentheses, court codes followed by number lists, and bare numbers in tables	Case numbers broken across lines may be missed. Check headers, table headers, and the opening of the document
Bank card	Checked	13- to 19-digit bank card numbers, judged by the check digit	Masked or truncated card numbers may be conservatively processed as a whole block. Compare digit by digit with the original to confirm unrelated digits were not replaced
Financial account	Checked	Bank account, passbook, and deposit receipt numbers under explicit fields, including partially masked, full-width, and space-containing forms	Bare digit strings without a field label may be missed. Check payment and receipt clauses
Online account	Checked	Online accounts and numbers under explicit fields	Nicknames or IDs without a label may be missed. Check registration and login information sections
Credit code	Checked	18-digit unified social credit codes, plus older organization codes with context such as "credit code"	Numbers without context that resemble bank card numbers may be classified differently. Check the entity information section
Registration or practice number	Checked	Numbers under explicit fields such as lawyer's practice certificate or practice license numbers	Abbreviated numbers or numbers written with only the last few digits may be missed. Check authorized-agent information
Property or vehicle number	Checked	Property and vehicle numbers such as property title numbers, real estate title numbers, VINs, and engine numbers	Abbreviated and parenthesized forms may be missed. Check property lists and enforcement information
Business number	Checked	Business numbers such as contract numbers, policy numbers, and project numbers	Custom-formatted numbers may be missed. Check the opening of contracts and attachment lists
Medical record number	Checked	Numbers under explicit fields such as admission or medical record numbers	The same digit string appearing outside the field is not replaced together. Check by field position
License plate	Checked	Mainland civil license plates (province abbreviation + issuing authority code + serial number)	Special plates such as military, police, and diplomatic plates may be missed. Check accident determinations and vehicle information
Date	Not checked	Various date formats, appearing as contextual information	When it overlaps with personal background (such as date of birth), only one candidate is kept. To replace it, check manually and confirm the scope
Amount	Not checked	Numerical amounts and currency formats	Amounts written in Chinese capital numerals may be missed. Check as needed, and review the claims and the judgment's operative part
Postal code	Not checked	6-digit postal codes under explicit fields	Bare digits are not treated as postal codes. Check as needed
Personal background	Not checked	Explicitly labeled background fields such as ethnicity, occupation, position, date of birth, and residence	Narrative background is easy to miss. When identity matters, check it and compare with date candidates
Health information	Not checked	Explicitly labeled injury, illness, and diagnosis fields	Narrative descriptions may be missed. Check medical and expert-assessment sections
Family relationships	Not checked	Explicitly labeled family relationship fields	Implicit relationships may be missed. Check as needed
Location information	Not checked	Explicitly labeled itinerary and track fields	Narrative locations may be missed. Check as needed

Categories will continue to expand with releases; which categories a particular version supports is subject to that version'srelease notes.

3. Replacement rules

Replacement text consists of a category word plus a letter, for example Name A, Phone A, ID A, Address A.
Organizations are replaced according to their name suffix with the matching category word, for example Bank A, Court A, Company A, Hospital A, Law Firm A; when the type cannot be determined, "Organization A" is used. This way the redacted document still shows what kind of organization it was.
When the same text appears several times in a document, the same replacement term is used, which makes cross-checking easier.
You decide whether to replace each item: items checked by default can be unchecked, and items not checked by default can be checked manually.
Replacement happens only at the selected locations; all other text and the layout are preserved according to the file structure, and the original copy is not rewritten.

Replacement is not recursive: a replacement term is not treated as source text again, avoiding chained changes.

4. Batch review

The same workspace can process a single document or a batch of documents; batch detection runs one document at a time in a queue.
Batch processing provides a summary review, and you can open any one of the documents for single-document review.
The same text in different documents is not necessarily the same subject. In a batch summary, identically named items must be judged one by one; do not treat them as an automatic determination.
If some files in a batch fail, the app continues with the remaining files and explains the succeeded, failed, and unprocessed items before export.

5. Originals and results

What is imported is a copy, and the redacted result is another copy; the original file is never rewritten, moved, or deleted. Undo is available only when a redacted copy exists; it retracts the redacted copy in the library and does not affect files already exported. Detection and review results are stored in the local index, so you can continue reviewing after reopening the app; if the original or the checked state changes, the old results must be regenerated.

6. Export and checks

Export the selected items for delivery; exporting an original prompts you first and allows cancellation.
Batch export does not overwrite existing files when names collide; if an export fails, the output from that batch is retracted and files that cannot be retracted are reported.
After export, open the files and check each one: confirm that the information you wanted removed has been handled and that the layout is not damaged.
Do not judge whether information has been deleted from the in-app preview alone.

7. Explicitly not supported

Scanned files and mixed image-and-text PDFs: outside full support.
Encrypted documents and pages without a text layer: rejected outright.
Cloud detection or cloud redaction: Wenyin does not do this and does not upload documents.
Sending documents for you: Wenyin will not send files on your behalf, and will not hand exported files to any AI service.
Accounts, subscriptions, and remote license revocation: not provided at this time.
Older Word formats (.doc), WPS, images, Markdown, and similar formats: not yet supported.

8. Known limits

Detection misses items and misidentifies them. Narrative private information, complex addresses, and unusual name formats still need manual additions.
Some PDFs and DOCX files are rejected: when the conditions for safe deletion are not met, the app prefers to refuse rather than generate a copy that might retain original text.
Hidden content and properties in DOCX, metadata and attachments in PDFs, and a full-structure check of the final file still require separate confirmation per the release notes.
For long documents over 20,000 characters, you can enable "Fast mode for long documents" in exchange for speed; the mode runs rule-based detection only, and the interface clearly marks that model detection was not used.
Apple silicon and macOS 12 or later are the current targets; Intel Macs and macOS 12 through 14 have not been validated (seeSupport and Updates).

Back to Wenyin home