Changelog

Questions in the interactive PDF editor are now detected by AI. The "Generate with AI" step on the toolbar takes a page range, the elements to place on the page and the difficulty, pool and topic defaults for the questions it will create; the editor stays read-only while the scan runs. What comes back is a draft: nothing is written to the question bank, every question is framed where it was read, and a preview button in its corner shows the text, the choices and the answer that would be saved. The report lists every detected question next to the answer it carries, and for books that do not print their answers on the page, the answer key is applied to the draft. Approving saves only the questions whose answer is known; the rest stay in the draft until an answer is supplied. Discarding a draft only ever discards what was never saved — nothing an approval wrote is undone — and the same question is never created twice across runs.

Paging and totals were fixed on the course, unit and lesson listings. The organization’s own content, the user’s subscription and the content granted through a package were each fetched separately and merged, so a page of ten could return thirty rows, page two was not the second page of the union, and the page count came out wrong. It hit exactly the roles that reach more than one source — organization admins, coordinators and tutors. A subscription dated in the future also hid only the course while leaving its units, lessons and items visible; the rule now applies to all of them.

The back office gained an "Agents" panel, with twin question generation as its first agent. You upload a question bank PDF (optionally with a page range) and pick the curriculum node the questions will hang from along with the generation and review models; the AI extracts the questions, writes copyright-safe twins of them, has a separate model review what it produced and repairs once whatever falls below the quality bar. The run executes on the server, so closing the tab does not affect it, and the panel follows it stage by stage. Generated questions are reviewed as preview cards — text, choices, difficulty and curriculum topic are all editable per question — and reach the question bank only once you approve them. A printable PDF can be produced as well.

Twin generation runs became reliable. Three runs can execute at once — a second run used to wait for the first to finish, which read as the panel being stuck — a run that goes silent is failed automatically so the panel offers Retry, and the same job handled twice can no longer overwrite a run that had already finished. Generation now uses the subject’s own glossary, so the terms a question tests are preserved verbatim while the choices and their order genuinely change. A single character the typesetter could not set no longer costs a whole run its PDF either.

Group lesson packages are grouped by subject. Package management takes a category per package (Sciences, Math & Computer Science, Social Sciences & History, Language & Literature, Arts, Engineering, Business & Economics, Other), and the pricing page shows a row of tabs for them under the group lessons tab. The row appears only when more than one category is present, and a leading "All" tab is selected by default, so uncategorised packages stay reachable.

A "Credit Ledger" page was added to the management menu. Every movement on student credits and the current balances are listed in two tabs, with search, filtering and paging all resolved on the server, and a balance row hands over to the ledger filtered to that credit. A movement opens in a right-hand sheet headlined by its minute delta, and shows the package that granted the minutes, the feature it came from, what was paid for that package and the payment behind it with its date, status and total. A credit granted by hand says so instead, since it has no payment record.

The tutor earnings list is aggregated on the server. The page used to pull every matching entry in 500-row requests and then group, filter, paginate and total them in the browser, with the name search and the subject filter applying only to what it had fetched. The list now renders one page from a single request, and the totals and the export are computed over the whole filtered range. Cancelled entries stay out of every earnings figure and a row with nothing left to pay is dropped rather than shown at zero — filtering by the cancelled status brings those rows back.

The payments and subscriptions management listings are scoped to the caller’s organization. The organization filter never reached the server, so the queries ran unscoped: an organization admin’s page usually failed to open, and when it did open, the counters above the list were counted over every organization. Platform administrators keep their cross-organization access.

Deleting an account now cancels its subscriptions. The subscription stayed active after the account was deleted, so the nightly renewal kept charging a card nobody can pay, then failed while addressing the cancellation e-mail to a user it could no longer find, and the same rows failed again the next night. The subscriptions the account pays for on someone else’s behalf — a guardian paying for a student, say — are cancelled with it, and the payment history is kept intact.

A "Metrics" page was added to the management menu. Six cards cover new users, active users, completed assessments, answered questions, payments with their revenue, and completed live sessions; each shows the value for the selected range, the change against the previous period of the same length, and a sparkline. Selecting a card drives the chart below it, which overlays the previous period as a dashed ghost line. Buckets follow the range — daily, weekly or 30-day — and every total is the sum of the buckets it is drawn from.

Elapsed exam time is recorded from the wall clock. It used to be tracked by a counter bumped a fixed amount per tick, and a background tab is throttled to about once a minute, so a student who switched away came back with part of the elapsed time never recorded. The timer on screen and the value stored now read the same clock, the time is written when the page is hidden or closed, and a checkpoint still in flight when the student finishes can no longer put the attempt back to "in progress".

The dialog that waits while open-ended answers are graded was fixed. A single failed request was read as "grading finished", so an attempt could be completed over answers that were never scored; a check that could not be made is now treated as unknown and polling continues. There is also a way out of a dialog that cannot be dismissed: if grading drags on, it offers to go to the results. Polling stops while the tab is hidden and checks immediately on return, and on the interactive PDF side a badge no longer stays stuck on "grading" after an error.

A student added to a class later is added to the class’s upcoming group sessions too. Participants were snapshotted from the class roster when the session was booked, so a student who joined afterwards had no participant row and was rejected at the join link. The missing rows are now filled in for the unfinished sessions of the class’s active tutor matches, and the join link is e-mailed and notified to the newly added student only. Joining by invitation code goes through the same path.

The comboboxes on the intro session form no longer show untranslated empty states, and their options load faster.

Date-specific availability is saved in one go. Editing a day dropped its existing slots first, which put the day back on its weekly hours for an instant, so an edit whose final state covered every booked session perfectly well could still be rejected: a tutor with an 11:00-14:30 override hosting an 11:00-12:30 session could not add 18:00-23:00 to that same day. Saving is now a single request in a single transaction, the coverage check runs against the resulting hours, and a rejected edit leaves the stored hours untouched. A failed save keeps the dialog open so the entered hours can be corrected.

The app pages got markedly faster. The sidebar read the current roles through a query of its own, which made a page rebuild itself two or three times per load and re-run every component and every query with it; roles are now read from what the session already carries. Translation files are shared with the request already in flight, and the translation watchers that piled up across client-side navigation are gone. The payments, subscriptions and earnings lists arrive in one query instead of an extra request per row, an exam checkpoint costs two requests a minute per student instead of six, and the grading-status poll reads a single marker rather than the whole queue. Requests per full page load went, for example, from 29 to 6 on the questions page, 18 to 6 on contents and 24 to 9 on users.

Fixed a fault that made screens asking for a user profile fail. The profile field could not be resolved on the server at all: it came back empty with an error recorded alongside it, and since the client rejects any response carrying an error, that surfaced as the page not opening rather than as a missing field.