Compare the writing style of documents and find which texts differ, on your own Windows PC. Your texts never leave the computer.
Windows 10/11 x64 · Offline · English / Russian / Polish
Load 4–200 documents in one language. Each should contain at least 300 words; about 1,000 is better. Linda Trace ranks documents by how much each differs in style from the rest, identifies the most different document, shows the features behind it (each value against the median of the other documents), highlights the 300-word fragments that differ most, and draws a 2D style map.
Load questioned texts and known-author samples: at least two samples per author and at least two authors. Linda Trace shows the nearest candidate authors by style. These are candidates, not probabilities.
Formats: TXT, DOCX, PDF, RTF, HTML and Markdown. TXT encodings UTF-8, UTF-16 and selectable Windows-1251/1250/1252. Export to HTML, PDF, JSON or a ZIP archive with file hashes. Saved datasets (cases) stay on your PC. Light, dark and high-contrast themes; interface in Russian, English and Polish.
The method combines about 250 stylometric features, character 3–4-gram vectors (a bundled resource of 2,000 features per language, trained only on documented public-domain and CC0 literary prose: 407 works), and a directional cosine-delta score into one ranking. Statistical p-values are not shown because the null model was not validated.
All analysis runs locally. Uploaded texts are never sent anywhere; the app only downloads update files (see Updates).
64.6 MB (64,618,674 bytes)
SHA-256: 619b97364bf6be825b20202472a438758b9d43e546be7c07d0f09b65b11b1453
70.2 MB (70,242,998 bytes)
SHA-256: 5303e60a9dbc3a7c4b6c6b35973eb8c9894334403c2f09721748651e206e788e
Per-user installation; no administrator rights; about 160 MB on disk; no GPU needed. Windows 10/11 64-bit. The app uses the WebView2 runtime, present on current Windows. The installer is not code-signed yet, so SmartScreen may warn: choose “More info”, then “Run anyway”, and compare the SHA-256 above. Data are stored in %LOCALAPPDATA%\ForensicAuthorship; uninstall asks whether to delete them.
The app’s “Updates” button checks a signed update list (Ed25519) published on Hugging Face and shows whether a new model pack or app version exists. Model packs install in seconds without reinstalling: SHA-256 is checked, changes are applied atomically, and the old pack is restored if installation fails. App updates come from GitHub Releases, are verified by signature and SHA-256, installed silently, and the app restarts. You can turn off the automatic check at startup in the Updates window. Only the update list and update files are downloaded; texts and documents are never sent.
Results show relative style distances within your corpus. They do not prove authorship, identity, authenticity or leaks, and do not replace expert examination. Short texts (under 300 words), lists, forms and slides are unreliable. Character n-grams can reflect topic or source as well as author. There is no unknown-author rejection with proven accuracy.
Independent held-out test on 92 historical literary works (authors who died 1852–1925): top-1 accuracy of the shipped fused ranking was 9/17 English, 9/19 Polish and 14/24 Russian. The plain n-gram score alone was 12/17, 12/19 and 14/24 respectively. On modern institutional prose (US Supreme Court opinions from 2024–2025, 5 justices × 4 opinions), top-1 was 4/10 for all methods. Modern and corporate texts are not covered by any claim. Samples are small; run your own checks before relying on results.
Proprietary build. Copyright 2026 project owner; all rights reserved. Use is under a separate agreement with the owner. You are responsible for rights to documents you analyse. Third-party component notices (THIRD_PARTY_NOTICES) ship inside the app.
Questions and bug reports: GitHub Issues.
