Legal
Documentation
Core Principle — Significance
Local Document Processing: PDF, image processing, OCR, recognition, redaction, pseudonymization, and local additional functions run on the device.
Multi-stage Recognition: Deterministic rules, context, local AI, local user learning, and document structure complement each other.
Secure Output: Redactions are implemented in the newly generated page; redacted plain text should not reappear via a search layer.
User Control: Uncertain findings remain verifiable; manual additions are possible.
1. What is Blursafe?
Blursafe is a desktop application for the secure editing of sensitive documents. The software helps to recognize, review, redact, or pseudonymize personal and confidential information.
The focus is on local processing. Document content does not need to be transmitted to an external document processing service for OCR, analysis, local AI recognition, local user learning, redaction, or pseudonymization.
Blursafe is designed as an assistance system: Automatic procedures support the review, while the professional decision regarding the specific need for protection remains with the user or the responsible person.
2. How does the typical workflow work?
Add documents: PDF, JPG/JPEG, and PNG files can be processed individually or together as a related collection.
Check sensitive content: Blursafe analyzes the documents using multiple local recognition procedures and provides recognized or uncertain information for review.
Redact or pseudonymize: Selected information can be removed or replaced by a substitute designation.
Check preview: Before the final output, the intended changes can be reviewed and manually supplemented.
Generate output: Blursafe creates a new PDF file and, depending on the workflow, can additionally generate a test report.
3. Which file formats are supported?
Native, digitally generated PDF documents
Scanned PDF documents
PDFs with a mixture of native and scanned pages
JPG/JPEG images
PNG images
Image files are treated internally like document pages. For mixed PDFs, each page can run through the appropriate local processing path depending on its structure.
The output is in PDF format. As far as text information can be sensibly preserved or newly generated, a searchable text layer can be used without reintroducing redacted original information invisibly into the output file.
4. Which processing steps run locally?
PDF rendering and page preparation
Image processing
Text recognition (OCR)
Analysis and recognition of sensitive details
Deterministic recognition and security rules
Optional local AI inference
Local user learning
Document structure and layout analysis
Redaction and pseudonymization
Creation of output files
Technical review of the generated output
No cloud OCR, cloud AI, or external document processing service is required for this document processing.
5. What data leaves the computer?
Document contents are not transmitted to external OCR, AI, or cloud services for the actual document processing.
Network functions are separated from this and can be used in particular for license management, software updates, or opening external purchase and information pages.
During license communication, technical license and program information may be transmitted, for example a hashed device identifier, license information, program versions, and the respective license transaction. Document contents are not part of this communication.
6. Does Blursafe require an internet connection?
For processing your documents: no.
OCR, document analysis, local AI inference, local user learning, redaction, pseudonymization, and PDF creation can take place locally. Product-related functions such as license activation, license verification, or updates may require separate network communication.
A failure of the internet connection does not lead to documents being outsourced to a cloud for processing.
7. Does Blursafe use artificial intelligence?
Blursafe can use an optional local AI component for additional recognition of sensitive details. The model, tokenizer, and runtime are used locally; document texts are not transmitted to an external AI provider for the actual inference.
The AI does not replace the deterministic security mechanisms. An AI hit is additional evidence and must not weaken existing strong protection decisions. If the local AI fails or is deactivated, Blursafe continues to work with the remaining local mechanisms. There is no cloud fallback for document analysis.
8. Is the local AI mandatory?
No. The local AI is an optional additional function. The basic product remains usable without this component. Activation is controlled and should not take place solely because an AI package is installed or sufficient RAM is available.
Specific hardware requirements may vary depending on the approved product status and target platform and are determined on the basis of practical measurements.
9. Does Blursafe learn from my decisions?
Blursafe can evaluate explicit user decisions locally and derive reusable learning rules from them. This user learning is not training or fine-tuning of an AI model.
Mere AI hits or automatic recognition results do not automatically become learning signals. Learning signals only arise from explicit user actions intended for this purpose. Repetitions within the same document are not simply counted as multiple independent experiences.
A learned positive rule can generate additional review evidence. It must not weaken existing strong or hard hits. Contradictory decisions are not simply resolved by majority rule; the learning effect in question can instead be paused.
10. What does local user learning store?
The permanent learning state is designed so that the learned personal value does not have to be stored as plain text in the Q-learning memory. Persisted data includes local identifiers, data type, opaque document identifiers, necessary state information, format information, and an installation-specific local random value.
The actual value can be processed locally for a short time during an explicit user decision in order to form an identifier from it. The local learning state can be deactivated, reset, or deleted.
11. Can local learning weaken protection decisions?
No. Local user learning is designed to be additive. Positive learning rules can trigger additional review, but cannot downgrade any existing hard or strong recognition. Negative learning information is not an automatic release filter and must not delete hits, force SAFE, or bypass a blocking final review.
12. How does the OCR work?
Pages without a usable native text layer – for example scans or image files – are processed with a locally executed OCR. The required OCR components and language data are provided locally.
German is the primary validated reference language. English documents can also be processed; individual deterministic recognition components may have a different coverage. For other languages, no general, full recognition promise should be made without separate validation.
13. What happens to rotated or photographed documents?
For image files, existing orientation information can be taken into account. For scan pages that appear suspiciously rotated, alternative 90°/270° alignments can be checked for the OCR.
An alternative alignment is not adopted solely because it recognizes more text. The selection is made conservatively so that a supposed OCR improvement does not come at the expense of already recognized sensitive information.
Strong perspective distortions or arbitrary photographic document corrections are not generally promised as automatically solvable.
14. How does Blursafe process multi-column and complex structured documents?
Blursafe can take the spatial structure of native PDF text layers into account. This is particularly relevant for multi-column pleadings, forms, tables, or label/value structures.
The goal is not to mix contents from different page areas just because they are at approximately the same horizontal level. The layout analysis uses existing local document geometry and remains separate from semantic recognition.
If no reliable structure can be determined, no uncertain geometric assignment should be forced. Geometric assumptions must not automatically become secure hits.
15. What should be considered with tables and forms?
For tables and forms, the spatial assignment between labels and values can be crucial. Blursafe treats this task as a structure problem and should not guess which fields belong together solely from the meaning of a text.
Uncertain spatial assignments are treated conservatively. Tables, multi-line cells, full-width headers, and staggered columns must not be unnecessarily degraded by layout analysis.
16. Does Blursafe recognize handwriting, signatures, stamps, and logos?
Such content can contain personal or confidential information, but is not equivalent to normal machine-readable text.
Depending on the released functional scope, Blursafe can automatically recognize visual sensitive content, mark it as requiring review, or make it accessible for manual additions. This may include, in particular, handwritten additions, signatures, stamps, company logos, and special graphic markings.
A general guarantee that every piece of handwriting, signature, or logo is reliably recognized automatically is not given. Uncertain visual findings must not go unnoticed and be treated as secure text decisions.
17. How does pseudonymization work?
Redaction and pseudonymization pursue different goals. Redaction removes information from the visible output. Pseudonymization replaces it with a substitute designation, for example "Michael Brandt" → "Person Alpha".
This can preserve the contextual connection of a document without the original designation remaining visible. Within a shared rule set, the same original designation can be consistently replaced by the same pseudonym. Projects or rule sets can be saved and reused.
18. What types of sensitive information can Blursafe recognize?
Personal names
Companies
Addresses
Dates
Email addresses
Telephone numbers
IBANs
File reference numbers
License plates
VAT identification numbers
Tax numbers
Commercial register numbers
Depending on the data type, different local methods are used. Not every category is recognized according to the same mechanism, and new categories can be further expanded with the product.
19. How secure is the redaction?
Blursafe pursues an approach in which sensitive areas are not merely covered by a movable black PDF object. The respective areas are altered in the image data during the generation of the output page.
This is intended to prevent the original visible content from simply remaining intact under a graphic cover. Any search layer generated must not reintroduce redacted original text into the file as invisible plain text.
20. Does Blursafe also check the finished output?
Blursafe can technically verify the generated output so that the security check is not limited exclusively to the original input text or the planned rules.
Such a final check can, in particular, re-examine the actually rendered output pages. This allows addressing cases where information was removed in the text layer but would still be visually present.
Even a technical review is not a promise of absolutely error-free recognition in every conceivable document form. Conspicuous or uncertain results remain in need of review.
21. Does Blursafe automatically recognize every piece of sensitive information?
No. No automatic procedure can guarantee finding every piece of sensitive information error-free in arbitrary documents. Scan quality, unusual fonts, complex layouts, OCR errors, language, handwriting, or graphic content can influence recognition.
Blursafe therefore combines several separate protection levels with manual control. Uncertainty should remain visible as uncertainty and not be output as certainty.
22. Can an AI or learning rule remove a secure hit?
No. Existing strong or hard deterministic hits have priority. Local AI and learning hints must not downgrade or remove these decisions. The direction of protection is intentionally asymmetric: auxiliary systems can trigger additional reviews, but cannot cancel protection decisions without a sufficient basis.
23. What happens if a local additional function fails?
Additional functions such as local AI or local user learning are not the sole basis of document processing. If a component fails, the failure should become visible while the remaining local and deterministic mechanisms continue to work.
In particular, a failure must not automatically lead to a document being considered secure. There is no silent cloud fallback for document analysis.
24. How is local learning information protected under Windows?
Local learning information is stored in the local user profile. Thus, the protection limits of the operating system profile generally apply. Other normal local users do not get normal access to the data of another user profile under a standard Windows profile boundary; administrators and SYSTEM can generally access local user data.
Blursafe should therefore not claim to cryptographically hide local learning information from an administrator of the computer. Additionally, the learning state is designed so that the learned personal value does not have to be persisted as plain text.
25. Does Blursafe support multiple user profiles?
A dedicated, complete multi-user or tenant profile model for local learning rules is not automatically equivalent to a Windows installation. Depending on the product configuration, a local learning context may exist per installation or user profile.
A further separation according to lawyer, clerk, client, or organization is an independent product and data protection decision and should only be promised if it is actually implemented and verified.
26. Is there a test report?
Depending on the workflow, Blursafe can generate a test report for the processing. This can, for example, document applied rules, processing steps, or hit counts in a comprehensible manner.
A test report does not replace the professional responsibility for the decision as to which information must be removed or pseudonymized in a specific process.
27. Can Blursafe process large volumes of documents?
Blursafe is designed for multi-document and batch processing. Multiple documents can be processed with shared rules or within a shared working state.
Specific performance limits and hardware requirements should be based on the respective released performance and hardware tests. A fixed maximum number of pages should only be published as a product promise if it is reproducibly proven.
28. Is there an API or automation possibilities?
Blursafe is primarily designed as a desktop application for direct use at the workplace. A publicly accessible general API or a general remote control CLI mode is not automatically promised.
Within the application, functions for recurring workflows may be available, such as batch processing as well as reusable projects and rule sets.
29. Can the generated documents subsequently be used in other systems?
Yes. Depending on the document and the processing carried out, generated PDFs can remain machine-readable or searchable.
As a result, pseudonymized documents can subsequently be used, for example, in document management, search, or other AI-supported workflows, provided this is permissible under data protection law and organizationally in the respective application case.
30. Is Blursafe GDPR-compliant?
A general statement like "Blursafe is automatically GDPR-compliant" would be too far-reaching. GDPR compliance depends not only on software, but on the specific processing operation, purpose, legal basis, and the technical and organizational measures of the controller.
Blursafe is technically designed for privacy-oriented document processing. This includes local document processing, local OCR, local AI, no cloud fallback for document analysis, no training of external AI models with customer documents, data-minimizing local learning information, secure redaction, and the separation between document processing and product-related network communication.
31. Is Blursafe GDPR-certified?
A statement like "GDPR-certified" should only be made if a corresponding reliable certification actually exists. Without such a basis, specific, technically verifiable statements are more meaningful, for example: "Document processing including OCR and local AI inference takes place on your device."
32. What should be checked before sharing an edited document?
The generated output should be checked before being shared externally. Special attention is advisable for unusual document layouts, poorly readable scans, handwriting, signatures, stamps, logos, graphic markings, or rare data types.
Blursafe supports this control through automatic recognition, local AI, local learning rules, layout analysis, manual supplement options, and technical verification. The professional decision regarding the specific need for protection remains with the user or controller.
33. What role does the user play despite automation?
A central one. Blursafe is designed to use automation without making it an opaque black-box decision. The user can review suggestions, mark additional areas, define pseudonyms, and correct decisions.
Uncertain AI, layout, image, or learning hints are not automatically raised to the highest protection level. This makes it understandable why information is being treated or presented for review.
34. Why does Blursafe combine multiple recognition methods?
Different sensitive details pose different technical problems. An IBAN can be recognized reliably in a different way than a standalone last name. A signature is not a normal text problem. A multi-column form is first a geometry problem before it can be decided which words belong together.
Blursafe therefore does not try to solve all tasks with a single AI model. Instead, specialized local mechanisms are combined, which can be checked separately and evaluated according to their evidence.
35. What exactly does "local" mean with Blursafe?
For document processing, "local" means: documents are not sent to an external document processing service for OCR, recognition, local AI inference, local learning processing, redaction, pseudonymization, or PDF creation.
Product-related functions such as licensing or updates can use network communication separated from this. The most precise statement is therefore: Your document content remains on your device during Blursafe document processing.
36. Summary of Security Principles
Document processing locally instead of cloud processing of content.
Multiple specialized recognition methods instead of a single black box.
Deterministic strong hits must not be weakened by auxiliary systems.
Local AI provides additional evidence, no cloud fallback.
Local user learning is based on explicit decisions and is not model training.
Complex layouts are treated as a geometry problem.
Visual content such as handwriting, signatures, stamps, or logos is treated separately from normal text.
Redactions are implemented in the newly generated output, not just visually covered.
The generated output can be technically verified.
Manual control and supplementation remain part of the workflow.
Note: This documentation is deliberately version-neutral. Specific product promises regarding individual functions, platforms, hardware requirements, and automatically recognized visual categories should always match the currently released product version and the associated acceptance tests.