Skip to content

What’s new

All notable changes to QualiLens. Release tags are semantic versions; each release also carries a build stamp (build YYYY.MM.DD-HHMM) that the in-app update check compares against your installation.

1.10.0 — 2026-10-09

  • Reports now list the model that actually answered each call, and indicate when a provider answered with a different model from the one the project chose. Each call records the model the provider reports, and the report header, the configuration section, and the audit appendix indicate any difference, in the app and in the Word report, with an amber notice in the app. A dated snapshot, such as OpenAI’s gpt-5.1-2025-11-13 for gpt-5.1, counts as the model itself. Calls logged before this release recorded no answering model, and the audit appendix counts them separately under the model requested.
  • gemini-3.7-flash has been removed from the wizard’s list of Google models. Google has deprecated it and now answers requests for it with gemini-3.8-flash at the same price, so the two choices had become one model. A project that already uses 3.7 keeps running on 3.8, and its report lists 3.8 for calls made after this update. For calls made earlier the report lists only 3.7, so a methods section should indicate that calls made since Google’s change ran on 3.8.
  • OpenAI’s default is now gpt-6.1-sol, and gpt-5.1 and gpt-5 have been removed from the list. OpenAI has deprecated both, and it shuts down gpt-5 on 11 December 2026 and gpt-5.1 on 1 April 2027. A project that already uses either keeps running until that date. The cost estimate prices gpt-6.1-sol at $2.00 and $10.00 per million tokens, gpt-4.1 and gpt-4o at their own prices, and gpt-5.1 and gpt-5 at $1.25 and $10.00 for projects that still use them.

1.9.0 — 2026-10-08

  • Other accounts on a shared computer can no longer reach your projects through the app. The page the app served handed its session token to whoever asked for it, and a connection to the app does not say which account on the computer made it, so on a shared Mac or a lab server another account could read your projects and use your keys. The launcher now opens a launch page readable by your account only, which signs your browser in, and the app gives the token to no one else. A tab opened by address in a browser that has not been signed in shows how to open QualiLens, and the launcher’s Terminal window names the launch page.
  • Word reports and checkpoint spreadsheets no longer fail on control characters. A form feed in a source, which PDF-to-text tools write at every page break, made the Word export and the checkpoint spreadsheet fail for the whole run. Such characters are now cleaned when a source is added and again when either file is written, which covers sources added before this release.
  • QualiLens handles untrusted files more cautiously. A .docx or a checkpoint workbook that would unpack to an implausible size is refused before it is opened, and ffmpeg may read a recording only from disk, never from the network.
  • QualiLens now uses pypdf 6.19.0, which fixes eight published flaws that let a crafted PDF make text extraction run for a very long time or exhaust memory.
  • Dependencies are pinned with hashes. Python packages, including the ones QualiLens uses indirectly, are installed at fixed versions and checked against the hash of the published file, and the launcher builds the interface of a GitHub clone with npm ci.
  • Intel Macs are not supported. A library QualiLens needs for encryption and for checking update signatures stopped publishing builds for Intel Macs, so installation there has failed since 1.5.0 with a misleading message. The launcher now says so plainly, and it tells someone running an Intel Python on an Apple-silicon Mac how to fix that.
  • Settings links to What’s New. Beside the version in the Application card, What’s new opens the website’s page of what each release changed, so the record is one click from the app.

1.8.2 — 2026-09-30

  • Gemini works again for new Google projects. Google now offers the Gemini 2.5 models only to projects that used them before and refuses them to others as “no longer available to new users”; the catalog offered only those models, so a new Google key failed at Test and at the first call of a run. The catalog now offers gemini-3.6-flash (the default), gemini-3.8-flash, gemini-3.7-flash, gemini-3.5-flash-lite, and gemini-3.1-pro-preview, with per-model prices; the wizard marks the last — paid tier only, since a free key has no quota for it. A project already set to a 2.5 model keeps it, and its estimate keeps 2.5 prices.
  • Check models catches a model Google lists but refuses. Google’s model list still includes the 2.5 models for projects that cannot use them, so the check reported them as live. For Google it now also asks about each catalog model through Google’s free token-counting endpoint, marks a refused model ✗ with Google’s reason, and the wizard disables it. Continue on the wizard’s model step runs the same check before it creates the project, whenever a key was just typed or no check has finished.
  • Gemini errors say what happened. A refused model says so and names the model Google suggests, instead of suggesting it was retired, and says when that suggestion is paid-only. A model with no free-tier quota — gemini-3.1-pro-preview on a free key — fails at once with the steps to Google’s paid tier, instead of after a minute of retries; a spent daily quota says when it resets; a paid project is pointed to a higher tier rather than to billing; and a per-minute limit waits as long as Google asks. An answer Gemini stopped for safety or recitation is refused as incomplete rather than parsed.
  • Gemini token counts include thinking. Gemini bills the tokens it spends thinking as output but reports them apart from the answer, and the Run screen, the report, and the audit log counted only the answer — a sixfold undercount on one call we measured. They now count both, and each call’s audit entry records the thinking share. Reports from earlier Gemini runs understate output tokens for the same reason.
  • Gemini 3 models run at Google’s default temperature, as Google advises, and the audit log says so; the older Gemini models keep the temperature QualiLens sets.
  • The wizard’s custom-model field no longer suggests Google’s speech, image, music, robotics, and agent models, which cannot do a coding pass.
  • The manual and the wizard describe Google’s free tier. Outside the EEA, Switzerland, and the UK, on a key whose project is not on Google’s paid tier, Google may use what you send to improve its products and human reviewers may read it, so it is unsuitable for confidential data.

1.8.1 — 2026-09-24

  • Check for updates now says what changed. When a newer release exists, the first paragraph of its release notes appears under the result in Settings, so the decision to install can be made there; the release page link carries the rest. The notes ride along with the check’s own answer, so the privacy claim is unchanged: one request to GitHub, only when you press the button.
  • The website has a What’s New page, written from this changelog, so what each release changed is readable without visiting GitHub.

1.8.0 — 2026-09-24

  • API keys are encrypted at rest. Each saved key is stored as ciphertext, and the secret that unlocks it is a small file outside the data folder — ~/Library/Application Support/QualiLens/secret.key on macOS, ~/.config/qualilens/secret.key on Linux and WSL, or wherever QUALILENS_SECRET_FILE points — created the first time it is needed and readable by your account only. A synced, backed-up, or copied database therefore carries no usable key. Keys saved by an earlier build are encrypted in place on the first start, and the app says so. A database opened on another computer, or after the secret was replaced, shows its keys as unreadable in Settings with the reason; paste them again or Remove them. Settings names the secret file beside the data folder and warns when either sits in a cloud-synced directory.
  • The data folder is readable by your account only, and so are the database and its sidecar files, set at every start. Deleted and overwritten content is zeroed inside the database rather than left in freed space, and Remove compacts the file, so a removed key leaves it rather than lingering.
  • The manual’s Getting Started, Troubleshooting, and Data, Privacy, and Governance chapters say what the encryption protects against and what it does not: the participant data remain plain text, and someone using your own account on your own computer can read both files.

1.7.1 — 2026-09-03

  • An installation updated to 1.7.0 by an older build showed only its build stamp: that build’s updater refused the new RELEASE file. The version now also rides inside the interface’s index.html, which every updater accepts, so the next update names its version whatever it came from. “Updates checked” says “today” rather than “0 days ago”.

1.7.0 — 2026-09-03

  • The version you are running shows in the footer of every page and on the Application card in Settings, with the build stamp beside it, and Check for updates names it as well. Bundles now carry a RELEASE file that package.sh writes from the changelog’s top entry.
  • A reminder to check for updates. When the build you are running is a month or more old and no check has been made in two weeks, the Projects page says so in one line, once per launch, with the path to the check; Dismiss hides it until the next start. It is worked out on your computer from the build’s date and makes no request, so the privacy claim is unchanged. The Application card in Settings records when you last checked.
  • A polish pass over the interface. Settings offers only the controls a state allows and keeps its policy text behind “How updates work” and “Moving the folder”; the project page names settings by the wizard’s own question labels, methods by name, dates without seconds, and sizes as words rather than “0k chars”; the report writes Yes and No where it wrote true and false; the wizard’s method cards say how many stages and reviews a method has instead of spelling out the pipeline; Delete on a project card is quiet until hovered. The content-analysis report no longer prints raw data structures into its prose, and its frequency figures appear once, in the table. A contrast pass on the palette: darker ink, cards lifted by a soft shadow, and the mark’s violet as the single accent for the primary action and the active step.
  • Settings is one screen: the providers as a single panel of rows, each with its status, its key hint, and its actions, and a key editor that opens in the row; Data and Application side by side as two compact panels of facts. Page titles are larger and section labels are small tracked eyebrows. The same structure carries to Projects (a panel of rows, each with a glyph for its method), the project page (runs and sources as rows, the configuration as facts), the wizard’s method cards (the glyphs), and the run page. A checkout that was never packaged reports the last released version with a plus sign, “1.6.3+”, beside its build stamp.

1.6.3 — 2026-09-03

  • Security review of 2026-09-03, five fixes. Checkpoint spreadsheets write every cell as text, so a name or quote beginning with = can no longer become a live formula when Excel opens the sheet. The local guard now requires the request’s origin to equal the app’s own origin exactly (scheme, host, and port), closing the route by which a page served by another program on this computer, on another port, could reach the API with the session cookie. Uploads stream to disk under a bound per kind (200 MB documents, 2 GB audio, 8 GB video) and document parsing runs off the server’s event loop, so a large PDF no longer freezes the interface. A signed bundle older than the installed build is refused as a rollback — the zip path asks first, the GitHub path never installs one. The release page link is shown only when it is a GitHub page of the QualiLens repository.

1.6.2 — 2026-09-02

  • The manual now shares the interface’s register and type: the same ground, ink, and spacing, Inter for headings and navigation, JetBrains Mono for commands, a serif for the text. The typefaces ship with the app under frontend/public/fonts/ with their licence (SIL OFL 1.1), noted in NOTICE.

1.6.1 — 2026-09-02

  • Settings: “Where your data live”.

1.6.0 — 2026-09-02

  • A checkpoint can be worked in a spreadsheet. Download as spreadsheet saves the code review (or the literature-synthesis extraction table) as an .xlsx workbook with a column for your action, one for a merge target, and one for notes; Upload spreadsheet reads it back and loads its decisions into the review screen for you to check before approving. Rows are matched by id, an unknown row is listed rather than guessed at, a code with no row is unchanged, and a workbook from another checkpoint is refused. Notes enter the audit trail beside the decisions they explain, and the workbook itself is kept with the run and named in the audit export. Definition boxes on the review screen now grow with their text.
  • A quieter interface: a cool near-white ground, one ink for text and the primary action, colour kept for done, waiting, and wrong. The app now ships its own type — Inter for the interface and JetBrains Mono for model ids, build stamps, and the audit log — so it looks the same on Mac, Windows, and Linux, and fetches nothing to do so. Figures follow the same palette; the thematic map’s connectors no longer cross code boxes.
  • After an update, the page waits for the new build and reconnects on its own once you have started the app again; nothing to close or reopen. A tab left open across any restart now reloads itself to pick up the new session, instead of showing a token error.
  • The launcher checks the port before any other work and, when it is taken, says what holds it: for a QualiLens server, the build it is running, when it started, and the exact command to stop it. A server left running in a forgotten terminal keeps serving the build it started with, so this is how you learn that the app in your browser is older than the folder. The served page and /api/meta now carry the running build.
  • Test runs no longer re-stamp the folder’s VERSION.

1.5.0 — 2026-09-02

Security and evidence-integrity release. Updating is recommended.

  • Security hardening of the local server and of the in-app updater. Update bundles are now signed, and only signed bundles install; an update is refused while a run is executing or awaiting review. A tab left open across an app restart now asks you to reload the page.
  • Excerpts state their own status: an excerpt whose quote cannot be found verbatim in its source is marked unverified in the report and listed outside quotation marks in the Word export; the audit appendix counts located and unverified excerpts. Quote location tolerates case, ligatures, soft hyphens, and PDF line-break hyphenation, and searches the segment the model was reading first, so a recurring phrase is highlighted where it was coded — on a 26-paper corpus this recovered about a quarter of the quotes that were previously discarded.
  • Literature synthesis: the extractor now separates a paper’s own findings from findings it attributes to other work (shown on the extraction review, not offered to the synthesis) and skips reference lists; the citation guard recognises names in any script, ignores stopwords from filename labels, and requires the year to match.
  • Reports now carry the configuration the run was frozen with, the models that answered, and a per-checkpoint summary of decisions; the audit trail exports as JSON from the run screen. A run freezes its provider, model, and setup answers at start, so a resumed or branched run keeps its model.
  • Thematic analysis defines and names themes before the theme review, so the names in the report are the ones you approved. Framework analysis charts a promoted emergent code across the sources before the matrix. Content analysis samples the opening, middle, and end of each source for the codebook, reports rates per 10,000 characters beside counts, and names the unit it counts. A confidence the model did not give is recorded as missing, not as 0.8. Findings narratives pass a quote guard.
  • Grouping of very large code sets runs in chunks and consolidates, instead of failing on the output limit. The coder is shown up to 300 codes for reuse (was 120).
  • QUALILENS_DATA_DIR moves projects, uploads, and keys out of the app folder; the app warns at startup and in Settings when the data folder sits inside a cloud-synced directory. The write-ahead log is folded into the database at stage boundaries and checkpoints.
  • Text ingestion no longer guesses UTF-16 without a byte-order mark, reads Word tables in document order with text boxes, refuses .rtf, and converts .aac for transcription when ffmpeg is present.
  • package.sh rebuilds the interface when its sources have changed (a fingerprint is stamped into the build and checked at startup and in the packaging test) and rebuilds the manual as part of packaging. Release 1.4.0 had shipped an interface built before its latest sources.

1.4.0 — 2026-08-31

  • After a successful update, Settings shows a full-screen confirmation page instead of a small notice that was easy to miss.
  • The launcher (run.sh) now checks disk space, verifies Node 18+, shows npm output during first-time setup, and detects a port already in use.
  • The QualiLens chevron logo now appears in the app’s navigation bar.
  • Bug fixes and internal improvements.

1.3.0 — 2026-08-30

  • The launcher checks disk space and Node 18+, shows npm output on first run, and detects a port already in use. Chevron logo. Manual updates; the release ZIP needs only Python. (Entry reconstructed from the working log; 1.1 and 1.2 were never published.)

1.0.0 — 2026-08-27

Initial public release.

  • Five analysis methods, each a staged pipeline with researcher checkpoints: grounded theory, reflexive thematic analysis, qualitative content analysis, framework/deductive coding, and corpus-grounded literature synthesis.
  • Researcher-led by construction: runs pause at review checkpoints; your edits are final and are not overwritten by later automated stages; decisions are written to an audit trail the report reproduces.
  • Provenance-first evidence: verbatim quotes with character offsets, a coded-source reader that draws the coding over each document in place, and page anchors for PDF sources.
  • Literature synthesis cites only from the uploaded corpus — never from memory. Synthesis support must reference located extraction quotes by id; ungrounded support is dropped and logged, and a citation guard scans the generated text for citation-shaped strings matching no uploaded paper.
  • Interactive report plus a formatted Word export, with method-appropriate figures that disclose truncation on their face — the grounded theory model drawn as a left-to-right paradigm flow.
  • Any review a run has passed can be revisited: a branch carries everything up to that review into a new run and reopens it, leaving the original run and its report untouched.
  • Four providers (Anthropic, OpenAI, Google, Mistral), an editable model catalog with a live model check, Whisper transcription for audio and video, cost estimation before a run, resumable runs that do not re-bill finished work, and honored cancellation.
  • The app runs locally; data go only to the provider you chose. In-place updates from a downloaded bundle or, pull-only and on demand, from this repository’s releases — the updater’s allowlist leaves your projects, keys, and uploads untouched.
  • A fifteen-plus-chapter user manual, shipped in the app.