Changelog

Work that has shipped to the internal test channel since 0.2.6 and has not yet gone out to every install.

• A correct chat answer on an ordinary table (one with no total or subtotal row) could be marked unverified by mistake. Fixed.
• Twelve everyday questions that asked about a group CLAIR could not actually match in the data (a value the column does not hold, a filter with no matching column) used to answer with the whole dataset's number instead. CLAIR now declines those and names the mismatch instead of returning a confident wrong answer.
• Chat, PDF, PowerPoint, HTML and CSV exports of a "show me the rows" answer now carry the actual rows, not just a sentence naming how many there are.
• A Full Dataset Report that could not build one of its sections now shows a distinct warning naming which section failed, instead of the same plain success message either way.
• A True/False (yes or no) column no longer blocks a What-if prediction or a batch-predict file that includes it.
• Signals' statistical correction now accounts for the full space of combinations it searched, not just the ones it kept, cutting false "confident" findings on data with no real pattern by roughly ten times at its most thorough search depth.
• A chart or dashboard panel that folds many rows into one number (a bar chart, a histogram) now says when it was built from a sample smaller than the full dataset, and warns when the table underneath carries a total or rollup row.
• An auto-generated dashboard now shows whether the AI model helped choose which columns to feature, or a plain automatic pick was used.
• When CLAIR's local database falls behind on saving, Settings now shows a plain-language notice with a "Try saving now" action, instead of no signal at all.
• The checks that catch a model leaning on a row-number column, or a threshold-derived flag that happens to match the outcome, now run the same way on every screen that shows what a model relied on, not only some of them.
• Settings was rebuilt around one consistent layout across every one of its tabs.
• The first-run setup wizard received an accessibility pass: clearer screen-reader labelling and a way out of a dead-end error screen.

Settings redesign, Predict support for true/false columns, and a round of chat trust fixes.

• Settings was reorganized behind one consistent card layout, replacing several different tab designs.
• Predict, Segments, Regression, Compare Groups and Experiments all accept a True/False column as an input or an outcome now, where before it was invisible to every one of them.
• A category column that Smart Encoding had already converted to a number is now usable as a Predict input or outcome, instead of staying invisible to the picker.
• A Deep Analysis or Predict result that never clearly beat the obvious guess (the majority answer, or the plain average) now says so plainly, instead of reporting a high-sounding accuracy number that hides it.
• Factor Finder's background ranking runs a second, more thorough pass over time and upgrades the result in place once it lands, the same way its Segments and Regression counterparts already did.
• Fixed a gap that could leave a fresh install without Predict or Advanced Statistics fully unlocked until the app was relaunched.

A large import-honesty pass on real messy files, Intel Arc GPU acceleration fixed, and several chat verification fixes.

• Rebuilt how CLAIR reads messy real-world files, tested against a corpus of genuine government data releases: the column delimiter is detected instead of assumed, a title banner above the real header is skipped rather than read as data, a header split across two rows is merged into one, and a row with more or fewer values than its header is flagged rather than silently mis-read.
• A row that looks like a total or subtotal is now flagged wherever it appears in a file, instead of only when it sits at the very bottom.
• Fixed several cases where a chat answer could be marked verified while quietly double-counting a total row, undercounting a repeated-measures dataset, or averaging over a rollup group as if it were a real category.
• A classification target whose categories are not already numbered 0, 1, 2, ... (a 1-9 rating, an age range) now trains correctly in Deep Analysis instead of failing with an unclear error.
• Intel Arc graphics acceleration was fixed: CLAIR now drives Intel's own graphics through Ollama's built-in Vulkan support, so an Arc GPU is actually used instead of silently falling back to the CPU.

The first broad field-test round: encoding fixes, first-run reliability, and early chat trust work.

• Fixed text encoding problems that affected files broadly across the first alpha builds.
• Machines with 8 GB of RAM now get a usable default AI model at first launch, instead of none.
• Fixed several first-run reliability issues in the embedding and model-download pipeline.
• Chat answers that list or filter rows are now checked against the data the same way aggregate answers already were, closing a gap where a listing could be wrong with no warning.
• Nine defects reported during this round's field test were fixed, several of which turned out to share a root cause.

A more reliable local AI engine and installer fixes.

• CLAIR now supervises the whole local AI engine process, not just its server, so a stuck engine is caught and restarted correctly.
• Fixed a startup race that could pick the wrong AI model, or the wrong process ending up on the AI engine's usual port.
• Fixed a slow cold start on some machines.
• Hardened the installer's setup steps and fixed an updater path that could restart CLAIR unnecessarily.

The version-numbering scheme used for the rest of the alpha was adopted here, alongside a local API authentication fix.

• The desktop app shell now authenticates every call it makes to CLAIR's own local backend, closing a gap where another local process could have called it directly.
• Startup timing is now recorded properly, so a slow launch can actually be diagnosed.
• Fixed a file-lock issue that could prevent a clean update on some machines.

The first installable alpha builds: install, sign in, upload a file, and ask a question.

• First alpha builds distributed to the invited cohort: a Windows installer, sign-in by emailed code, and the core workflow of importing a spreadsheet and asking questions about it in Chat.
• Fixed an uninstall bug and made a failed cleanup step report honestly instead of claiming success.
• Chat moved out of an early-access label into the app's primary way of asking questions.
• One-click in-app updates: CLAIR shows an "Update & Restart" banner when a new build is available, with no re-download or re-install needed.