Linda updates

• A pre-check guide recommends at least 300 words of continuous prose. Very short texts and lists, rules, formulas, forms and slide-like material may not be assessed reliably. This is a recommendation, not a quality guarantee; “no signs found” does not prove human authorship.
• PPT/PPTX files are unsupported. Exporting slides to PDF does not make slide content suitable for analysis; text PDFs are supported.
• File replacement, model-folder moves and cancellation are safer and clearer. Diagnostic archives redact local profile paths.
• Microsoft Store package 2.0.6002.0 is prepared for submission and will become available after submission and review; the GitHub installer release is 2.0.6.2. Models remain version 1.3.4; no detection-accuracy improvement is claimed.

• Assay uses FP32 stable DirectML on Windows to reduce CPU/GPU score drift; Loupe remains on CPU. GPU memory use is higher than FP16; 4 GB compatibility is not certified.
• Assay samples the same context anchors on CPU/GPU.
• Model identity, independent Essay/Multi evidence, Unicode offsets and document extraction are corrected.
• Selected highlighting, counters, word shares and exports stay synchronized. Document confidence and local coverage are separate.
• Weights remain 1.3.4. Short AI inserts and edited texts may still be missed; these fixes do not certify detection accuracy.

• First-run tips now use the interface language (they were always in Russian).
• Reports (PDF and HTML) show the Linda mark in the header; the short PDF report no longer prints the old name.
• The installer shows the correct Linda Loupe size, about 0.3 GB.
• Models, verdicts and licence keys are unchanged.

• Linda is now the name of the product. The full ensemble is now called Linda Assay, the fast version is Linda Loupe.
• Models, verdicts, settings and licence keys are unchanged; keys work as before.
• New Metro-style logo: a tile with three lines of text, part of the middle one highlighted, just like the AI marking in the app. New icon for the app, the installer and the Start menu.
• The logo in the app header follows the light/dark theme.

• Linda-Pro Lite takes the token scores for highlighting from the same model passes that compute the verdict and reads only the rest of the text: highlighting is about 2x faster (our measurement: 2.1 -> 1.0 s per 1,000 words on 4 threads) with no loss of highlighting accuracy (85% of words labelled correctly on mixed texts, false AI mark on 3% of human words).
• Verdicts are unchanged.

• Sentence highlighting is now computed from tokens in the same model pass as the verdict instead of repeated passes over pieces of the text. Linda-Pro Lite highlights a 3,800-word text in about 4-6 s on a modern processor, without the huge uniform blocks of 2.0.4.x.
• Highlighting accuracy on mixed human + AI texts with a known boundary (our measurement, EN/RU/PL): Linda-Pro labels 90% of words correctly (2.0.4.x: 90% at the cost of many repeated passes), Linda-Pro Lite 88% (2.0.4.x: 67%); about 3% of human words get a false "AI" mark.
• The verdict and its accuracy are unchanged: same models, same thresholds.
• If the document as a whole is "unclear" but AI is marked in almost all of the text, the Authorship box says "AI"; if nothing is marked it says "unclear" instead of "mixed".
• Text scrolling no longer stops working after the model set is changed.
• History and document comparison open faster.

• Linda-Pro Lite: sentence highlighting is fixed. Before, almost the whole text was coloured "possibly AI" for human and AI texts alike, and no sentence was ever marked "AI". Now about 9 in 10 words of AI texts are marked "AI", and 0-3% of words in human texts.
• The verdict no longer contradicts the highlighting: if the app says "written by a human" while a noticeable part of the text (a quarter of the words or more) is marked AI, the verdict becomes "unclear".
• Linda-Pro Lite is more careful: when the Lite model sees AI above its own threshold but the combined score is lower, the verdict is "unclear" instead of "human" (an AI text edited by hand is no longer passed off as human). In the "Both sets" mode this also applies to the Lite verdict.
• The "possibly AI" band of Lite highlighting is wider and starts where the "unclear" zone of the verdict starts.

• Closing the window no longer leaves the program hanging in the Task Manager; a new launch ends an old hung copy by itself and no longer opens its old interface.
• Loading the models into memory at start no longer blocks the restart for an update ("finish the check" with no check running) or the applying of new models.
• Error messages of the update buttons stay on screen longer.

• New, improved Linda-Pro Lite (about 0.3 GB): retrained on texts rewritten by stealth humanizers, without Llama; it finds disguised AI text better and runs on any processor.
• New "Both sets" mode (Settings > Models): next to the Linda-Pro verdict the app shows the Lite verdict for comparison.
• With only Linda-Pro Lite installed, the interface shows only it: no rows or buttons for the Essay and Multi models; model names in reports (PDF, Markdown) match the installed set.
• Choice of set in the installer and in Settings; the graphics-card choice is kept.

• Two model sets: Linda-Pro (the most accurate, best with a graphics card) and Linda-Pro Lite (about 0.3 GB, runs on any processor, a 3,800-word text in seconds). Choose either one or both during installation and later in Settings ("Models").
• Lite finds noticeably fewer AI texts after "humanizers" than Linda-Pro (10-14 points lower in our measurements, much lower on deliberately disguised texts); use Linda-Pro for an important text.
• On a computer without a graphics card the app uses Lite by itself. Models download automatically.

• After new models are downloaded the app now offers “Restart” instead of showing the “Update models” button again.
• An old failed-install message (“Installer exit code 5”) is no longer shown at every start once that version is installed.
• Rows of the batch table no longer flicker while several files are being checked.
• Short text (under 150 words): a clear warning above the result that the estimate may be inaccurate.
• The results panel can be widened or narrowed by dragging the divider between the text and the results (double-click resets it).
• The built-in downloads flyout no longer covers the buttons at the top right.

• Updates no longer fail if you open the app during installation: it shows a message and opens by itself when the update is done.
• Launching the app again while its window is open brings that window to the front instead of opening a browser.

• Model weights 1.3.1 (the app updates them automatically, no new installer): the training data no longer contains texts generated by Llama models; instead we added permissively licensed human learner essays (ASAP 2.0, IELTS sets) and rewrites from open models.
• English: HumanizerBench 81.0% (was 80.8%), Oct 2026 cycle 84.0% (81.3%), three unseen generators 95.5% (94.6%), Chicago after StealthGPT 88.0% (86.6%). False flags: TOEFL 5.5% (6.6%), W&I 0.4%, ELLIPSE 0.4%.
• Russian: 80% of AI texts at 0.6% false flags (was 78%); Gemini-rewritten 59% (50%). Polish: 79% at 1.4% false flags (was 80% at 0.9%); Gemini-rewritten 56% (48%).
• Honest limits: false flags on MAGE human texts rose from 0.9% to 2.0%, and on the Polish set from 0.9% to 1.4% (5 of 350 texts instead of 3).

Technical report: https://doi.org/10.5281/zenodo.23242014

• Tabs: keep several texts side by side (up to 8, the + button or Ctrl+T). Each tab has its own text, file and result, and switching loses nothing.
• A check belongs to its own tab: while it runs you can type in another one, and when it finishes the tab shows a green dot. One check runs at a time.
• A file from Explorer (“Check in Linda-Pro”) opens in a new tab and does not overwrite the current text.
• Closing a tab whose check is running cancels that check.

• Automatic model updates: when a release brings new models they download in the background, are verified and replace the old ones (old files and caches are removed). On by default, can be turned off in Settings.
• Protection against wrong input: texts in unsupported languages (Arabic, Chinese, Greek and others) get a clear message instead of a false verdict; a binary file renamed .txt is rejected; file errors are explained in plain words; Windows-1251 and UTF-16 files are read correctly; German and Ukrainian texts show a warning.
• A one-page PDF summary for a teacher or editor: the verdict, what it means, the figures, the most suspicious sentences and a note that it is not proof.
• 70 “silly” scenarios were tested (empty and huge texts, odd files, wrong paths, rapid clicks) with no failures.

• Simple mode by default: one screen, a verdict in plain words, warnings for unreliable cases (short text, mixed languages, many numbers) and “Most similar to AI”; the Detailed button brings all panels back.
• A 3-step tour on first start (skippable; “Show the tips again” is in Settings).
• Models folder: choose it in the installer, on first start and in Settings, with moving of the models already downloaded. Free space is checked before a download and a time estimate is shown.
• The first check no longer waits: models are prepared in the background right after start; the header shows what computes; Settings has Check speed.
• “Check in Linda-Pro” in the Explorer menu for .docx, .pdf, .txt, .md files (a checkbox in the installer), a Check clipboard button, and files can be dropped anywhere in the window.
• Save diagnostics for a support e-mail: versions, graphics cards, logs, no checked texts.
• 150% font, hints for terms, fixed light and high-contrast themes, the arrow of the Advanced block.

• The settings window no longer hides under the top bar.
• Automatic graphics-card choice is more reliable: a discrete card is recognised by its video memory and the most powerful one comes first.
• Progress bars for file upload and reading, analysis, batch checks and document comparison.
• The interface language switches with RU / PL / EN buttons in the header; settings got a gear icon; theme, font size and device choice moved to the left panel.
• Analysis mode and model choice moved into a collapsible Advanced block, so the default screen is simpler.
• The in-app change list shows only version numbers and titles with a link to the full list.
• The Check for updates button moved to the updates section.

• New 1.3 models (Essay and Multi, each merged from three trained variants) for all three languages. English now uses the same routed path as Russian and Polish, on shared models: switching language no longer reloads them.
• Polish: on the main test set detection of AI text rose from 39% to 80% at 0.9% false flags (was 0.3%); Russian: 57% to 78% at the same 0.6% false flags; English: recent-model and humanized texts are caught noticeably more often (HumanizerBench 70% to 81%).
• Honest limits: in English, false flags rose slightly on non-native writing (TOEFL exam essays: 4.4% to 6.6% on 91 texts) and on lightly edited human text. In Precise mode the same TOEFL essays give 1.1% false flags (1 of 91), at the cost of catching fewer English AI texts (HumanizerBench 69%). Use it for schools and universities.
• The model package is about 4 GB; peak GPU memory is about 1.6 GB. The update does not delete old model folders.

The verdict threshold was frozen on development data; the website test sets were not used for selection.

• PDF reports now use the bundled renderer without Edge, Chrome or Brave. One A4 layout includes vector charts, structure map, metrics, annotated text and Unicode for English, Russian and Polish. Silent switching to the old exporter has been removed.
• PDF/DOCX parsing runs in a separate Windows LPAC process with restricted file and network access. Limits remain: 20 seconds, 512 MiB, 20 MB input, 4 MB output and 500 PDF pages; DOCX entry count and expansion are bounded.
• Current-user Windows DPAPI encrypts history. Existing records migrate into protected storage; writes are atomic and deletion respects retention policy.
• Restart and page-refresh drafts also use DPAPI. New text is no longer stored in plaintext sessionStorage; legacy drafts migrate while preserving text.
• The audit ledger uses an HMAC chain, independent checkpoint and recovery of interrupted writes. Manual clearing is disabled; organization retention policy controls storage.
• Fixed the blank WebView2 window after minimizing: the surface repaints without reloading the page or losing input. The dark window background is retained.
• An explicit PerMonitorV2 DPI manifest is included, with Windows long-path support for the temporary document worker.
• DirectML checks the actual execution provider. Switching CPU/GPU releases previous sessions and refreshes device status.
• Fixed isolated document-worker preparation and error handling: damaged documents do not block subsequent work or expose internal paths to users.
• Background update download, hidden installation, automatic relaunch and unsaved-text recovery remain supported. The stable channel serves the normal Windows installer separately from Dev.

Model weights, calibrations and commercial terms are unchanged.

• Background update download, hidden installation and automatic app relaunch; fixed the helper waiting indefinitely after a successful install.
• Installer failure reopens the previous app, records the error and allows retry; failed packages are not automatically relaunched.
• Unsaved text is restored after updating; temporary drafts are acknowledged and removed, respecting storage policies.
• Duplicate update clicks are blocked, notes match the offered version, and restart is blocked during analysis, queued work or document parsing.
• Package signature, size, version and checksum are verified. Failed model updates roll back changed, added and deleted files with the manifest; fixed open-file handling on Windows.
• Cancellation belongs to a job and tab, covering early cancellation, page reload and disconnect. Cancelled results are not stored in history.
• PDF/DOCX use a separate process: 20-second timeout, 512 MiB memory cap, 20 MB input, 4 MB output, PDF up to 500 pages; DOCX entry and expansion limits. Parsing keeps HTTP responsive.
• Late uploads cannot overwrite newer input; damaged documents fail without blocking subsequent work.
• Legacy single mode executes only the selected model. CPU voters load on demand, with optional preload.
• Single verdict/probability agree with history and JSON/CSV/HTML exports, including model and analysis scope.
• Stylo without sentence evidence no longer displays NaN, fabricated authorship shares or a human-authorship claim. Missing scores and text analytics without labels are handled correctly.
• Uncertain sentences are not exported as explicit AI signals. Missing highlighting does not establish human authorship.
• Precise mode requires the full ensemble; UI resets incompatible choices and API rejects them before model loading. Routed calibration supports the full ensemble.
• Editing text, mode or model invalidates stale results. Fixed human-verdict confidence, zero scores, history percentages and opening history records.
• Atomic settings; Origin/Host and iframe protections; CSV formula escaping; storage/export/folder/licence policies respected.
• Metro UI: flat colours, square corners, keyboard focus and reduced motion. History / Batch / Compare centred on the window, using a separate row on narrow windows.
• Status polling has a timeout; translations no longer redraw continually; update errors avoid blocking dialogs.
• Website mobile menu closes on Escape, navigation and outside clicks; full release notes added.

Weights and calibrations are unchanged.

2.0.1 fixed the export buttons (PDF, HTML, CSV, JSON, Markdown), which did nothing in the app window; 2.0.2 fixes the in-app update: after the download a Restart and update button appears and the app installs the new version by itself. If you have 2.0.0 or 2.0.1, the update button in those versions cannot install anything: please download Linda-Setup.exe from the latest GitHub release once and run it over your current copy (history, settings and models are kept). From 2.0.2 on, updates work from inside the app.

What is new: a PDF report with charts (probability along the text, authorship shares, where the text sits against human and AI samples, structure map), plus HTML, CSV and JSON exports; updates download in the background and apply after a restart, with rollback if something goes wrong; the interface in Russian, Polish and English; appearance settings (theme, accent, density, font size, corners); for organizations: batch check of a whole folder, an audit log and policies in enterprise.json. The word and character counter is fixed. The models are unchanged (version 1.1). Version 1.x updates itself from inside the app.

Fix for version 1.2: on polished AI texts the verdict said AI while the sentences were all coloured as human, because one sentence is too short for the models on its own. Now every sentence is scored alone and in its context (about 120 words around it) and the higher score counts, so a whole AI text is shown as AI and a lone strongly-AI sentence still stands out. On our test texts 64-78% of the sentences of AI texts are marked AI and another 18-23% possibly AI, against 0% AI and 2% possibly-AI for human texts. The thresholds for AI and possibly-AI now come from the program itself, so the main window and the full reader always show the same colours. The models are unchanged. Versions 1.1.x and 1.2.0 update themselves from inside the app.

What is new: (1) History kept only on your computer: every check is saved with its text and result; folders (drag a row onto a folder), search, sorting, bulk delete, automatic clean-up after 30, 90 or 365 days, switch it off any time. (2) Batch check: drop any number of PDF, DOCX, TXT or MD files and they are checked one after another, with CSV export. (3) Document comparison: two versions side by side with the passages they share (8+ identical words) and how the AI colouring changed; or an overlap matrix across many documents. (4) Full reader: open any saved check or compared document in the full analysis view and go back to the list. (5) Authorship verdict: human, mixed or AI, with shares of AI, possibly-AI and human text. (6) Text analytics and structure: sentence rhythm, lexical diversity, stock phrases, repeated phrases, an AI-probability strip along the text, a paragraph map and a violin plot of sentence scores against human and AI references. (7) Every sentence is scored on its own again (single AI sentences among human ones are visible), the possibly-AI band starts lower, and flat Metro animations. Everything still runs on your PC; nothing is uploaded. Limits: sentence-level detection is about 70-75% accurate, weaker than whole-text detection, and the structure view describes a text but is not part of the verdict. Version 1.1.x installs update themselves from inside the app.

The app now runs the models on any DirectX 12 graphics card through DirectML, with no CUDA or ROCm installation: 8-12 times faster than the CPU on our test PC (AMD Radeon RX 9070 XT: 0.3 s for 1,000 words instead of 3.5 s), identical verdicts. Automatic mode uses a discrete card; the first start prepares the GPU models once (about a minute). Choosing GPU on a PC without a usable card no longer breaks the app: it falls back to the CPU.

The colour map now follows sliding windows of about 300 words, so the boundary between human and AI-written passages is accurate to roughly 100 words, and the Multi view no longer stays green. The models can load right at start-up (Settings; turn it off and you get a warning that the first analysis will be slower). New Metro-style look for the app and the online demo. Licence keys are tied to the computer they are activated on and can be moved with Remove key; complimentary keys work offline. Version 1.1.0 installs update themselves from inside the app.

The three detectors were retrained on real outputs of AI humanizer tools. On the public HumanizerBench set (not used for training) the share of humanized texts flagged as AI rose from 31% to 70%; false flags on TOEFL exam essays fell from 8.8% to 4.4%. The technical report (DOI 10.5281/zenodo.23261670) lists the data, tests and caveats.

Data, evaluation and limitations of Linda-Pro 1.1 are described in an open technical report on Zenodo.