Status: direction accepted for implementation, 2026-09-22. Companion to the delivery plan; teacher pilot validation remains pending. The interaction mockup uses synthetic content and demonstrates review decisions; it is not a working importer or evidence of recognition quality.

R4 (v0.10.0) rebuilt every import screen on the design deck, under /teacher/imports, and added the pasted-text source. §§1 to 11 are the direction of 2026-09-22; §12 records the screens as built and wins where it differs: the route table in §2 is corrected in place, and §12.7 lists what the deck drew that the product keeps, what it gates and why.

Use Quizzivy's current zinc/charcoal components, divided settings-style rows, restrained borders/shadows and clear action hierarchy. The reference dashboard's organization informs spacing and navigation, not decorative charts or extra KPIs. The workspace uses the available width for comparison. Reading line length is bounded inside each pane, rather than leaving an unused third of the page.

The default experience asks the teacher to resolve meaningful decisions. Technical details, parser/model versions and retry diagnostics belong in an optional processing-history panel. No confidence percentage, “AI magic” decoration, fake progress, automatic answer invention or blanket “accept all errors” action.

Vietnamese first; every shipped label/help/error has an English key. All status states have text as well as icons/color. Use “Cần xử lý”, “Cần xác nhận”, “Thông tin”, “Đã lưu”, “Chưa lưu được”, and “Đáp án chưa có trong tài liệu”. A file that finished processing says “Sẵn sàng rà soát”, not “Nhập thành công”.

Route	Purpose	Entry / primary action
/teacher/imports	Searchable import history	Secondary entry from tests; "Import from Word/PDF" and "Paste a test"
/teacher/imports/new	Upload or paste, optional key	Tests list split action beside "Tạo đề"; "Bắt đầu xử lý"; ?source=paste opens the paste mode
/teacher/imports/:id	Current processing state or resumable import	"Tiếp tục rà soát" when ready; committed state links to resulting test
/teacher/imports/:id/review	Source-linked correction workspace	"Xem trước & hoàn tất"
/teacher/imports/:id/confirm	Learner preview and validated summary	"Tạo bản nháp đề"

There is no /compare/:runId route (WU-05): a newer machine draft that arrives while the teacher has edits is offered inside the review, beside the teacher's version, and applied with adoptWordImportReprocessed (§12.5). The old /admin/imports* paths redirect to these (T-R4.5).

Do not add many permanent global sidebar items. The tests area exposes imports and history clearly; the dedicated review route can collapse global navigation while preserving a labeled way back. A direct URL must resolve to a helpful state for processing, failed, cancelled, stale or committed imports.

These entries follow GET /teacher/imports/capabilities:

• Intake off. The tests list shows neither entry, and every /teacher/imports route explains that import is not enabled (ImportsGate).
• Intake on, processing off. The history, review and commit stay available. Upload, retry, reprocess and "start over" are replaced by a calm note saying why. A queued import says it waits for processing to return.

Imports create a new test initially. Starting from an existing builder flushes the existing draft, then clearly states that import creates a separate draft; it never silently replaces an open exam. Append/merge into an existing test is not implicitly added to this milestone.

Columns: filename/title, creator, created time, status, unresolved issue counts, resulting test and row actions. Search and status filters live in the URL and survive review/back navigation. A row's primary action follows its state: processing → view progress; needs review → continue; failed → view cause/retry; committed → open test. Do not auto-start a second import from the same row.

“Tạo đề khác từ tệp này” is a distinct action explaining that it creates a new import identity. Retrying resumes the current job policy. Bulk removal, where available, lists eligible items and per-row failures; active/committing/referenced items cannot be swept away. Archive/delete follows the approved retention policy (17-word-import.md §1.38). The upload page states it, and a reviewable import warns that it closes after the idle period. Once files are removed, the detail and review pages say so calmly, without download or review links.

Empty states differ: no imports offers upload; no filtered results offers clearing filters; unavailable service preserves controls and offers retry. History remains useful when the AI provider is disabled.

Use a page, not a long modal. The essential fields are visible without an accordion: exam file; optional answer-key file; optional audio attachment/library selection. Show filename, actual validation status, size, role and remove/replace action. Drag-and-drop has an equivalent labeled file picker. Input limits appear before upload and authoritative server errors appear beside the affected file.

Recognition: “Tự nhận diện” by default; “Dùng cấu hình đã lưu” and a downloadable optional template remain secondary. Advanced settings cover only explicit source conventions: numbering restarts, answer markings and score defaults. A profile must show a concrete example and its scope before applying; never expose a raw prompt editor to teachers.

If AI processing is configured, show where processing occurs and what data is sent in plain language. This uses an admin-managed approved configuration. Teachers do not choose infrastructure/model knobs on every import. Private mode must not silently use cloud when unavailable.

"Read with AI" (plan 17 §1.42, D4) is that case, and D19 decided it on 2026-10-10: cloud processing is approved, the admin-managed configuration is the server's environment and ai_settings (plan 17 §8), and the page says what is sent. The Sources pane of an import that asked for it says that scanned or photographed pages are read by an AI service outside Quizzivy and that pages with selectable text are not sent. The review reports what left the system as an informational note, "{n} pages were read by an AI service outside Quizzivy", and a page the service could not read as a note that asks for a decision. A scan that is refused names why and what to do: a file over 40 pages, no readable page, too few credits (nothing was sent and none was used) or a paused teacher. The card that turns it on is the Imports task T-D4.39.

Preflight handles password protection, mismatched extension, corruption, macros, scans and limits distinctly. The exam and the key may each be .docx or a PDF with a text layer; the limits line says a scanned PDF is refused unless the import is read with AI. A PDF that turns out to be a scan, locked, broken or too long fails processing with its own message and asks for another file (plan 17 §1.39). Replacing a rejected file preserves the other valid inputs. A partially uploaded companion key never appears as fully received.

The same page also takes a test pasted as text in place of the exam file. A "Source" switch chooses between uploading a file and pasting. The paste mode has a plain text box with a character count against the server's limit, a quick count of what was found so far, and no answer-key slot. The layout is in §12.3; the copy and tests are in 74-r4.md T-R4.57, the source and its recognition rules in T-R4.55 and T-R4.56, and the scope and what was built in 17-word-import.md §1.41.

What the teacher is told holds on every screen:

• Plain text only. No screen says formatting was kept, because none is. The source is named by a localized "Pasted text" label taken from its format, never from a stored file name.
• The page states the conventions the recognizer reads, and no others: put a heading on each part, number each question, put each option on its own line, and mark the answer with * after the correct option, an "Answer: B" line or one "Answer key: 1-B 2-C" line.
• The quick count follows the server's rules for text sources, so it cannot contradict the review. It is a hint; the full check runs in processing.
• Later screens name the text, not a file. Processing checks the text rather than the files, review shows the one source as pasted, and the original downloads as a .txt. A failed pasted import offers a new import, which opens the paste mode again.

Compact real stages: “Kiểm tra tệp” → “Đọc nội dung” → “Nhận diện cấu trúc” → “Đối chiếu đáp án” → “Kiểm tra đề”. Show elapsed time and current stage. Use a percentage only if the stage has a measurable denominator; otherwise use a plain activity indicator, without an invented time remaining.

Copy states that the teacher can leave and return from history. Closing the tab does not cancel a run. “Hủy xử lý” is an explicit action; explain that saved source and review data remain according to retention. A cancellation racing with commit shows the authoritative result rather than claiming cancellation unconditionally.

A failure card identifies what failed, retained progress and a specific action: replace an invalid file, retry a transient service failure, or return later when the provider is disabled. Processing history is expandable and avoids raw document text or stack traces. Completion opens review on an explicit action, preserving the teacher's current work if processing finished in another tab.

Top: filename/test name, concise “X câu · Y phần”, save state and next-step action. Below: section/group selector, issue filter and previous/next finding. Main space has two adjustable panes: source (~42%) and reconstructed editable exam (~58%). Issue details are inline with the selected field. Do not permanently stack a global sidebar, outline rail, source pane, editor and findings sidebar together.

At wide desktop widths, an optional outline can open in the shared side-column mechanism; it is collapsed by default if comparison would become cramped. Use existing resize controls, visible focus and keyboard resizing. Persist widths separately from other page panels.

Source pane: exam/key selector, “Nội dung trích xuất” / “Bản xem trang” when an actual rendition exists, and a labeled original download. Clicking a question or finding locates and highlights the corresponding source. Clicking a source block reveals its mapped item(s), metadata use or exclusion. Do not force synchronized scroll while the teacher is independently reading another passage.

Source highlights cite stable block/run/table-cell references. Show page numbers only from actual rendering and make conversion/font differences visible. When exact text-to-page highlighting is unavailable, select the known page/region and retain precise block evidence; do not fabricate a bounding box.

The selected card includes original printed label, normalized position, type, points, prompt, options/blanks, answer controls and evidence. A small toolbar supports the approved formatting marks, table actions and asset insertion. Question type changes warn about affected answer data and preserve source evidence. No heavy rich-editor instance is mounted for every question in a large exam.

Use visibly different labels for explicit source data, inferred structure, defaults and teacher edits. Reveal detailed provenance on demand. The teacher can edit a value directly, but the UI never describes that edit as an extracted source fact. Choice options keep stable local IDs when reordered, so keys follow the option.

Question/group actions: split, merge, move, exclude with reason, restore exclusion, and attach/detach shared material with dependency checks. Dragging has explicit “Chuyển đến nhóm” and move up/down alternatives. Splitting/merging previews the resulting boundaries, answer associations and points before applying.

Finding	What the teacher sees	Concrete resolution
Missing answer	“Tài liệu chưa có đáp án cho câu này” and searched evidence	Enter explicit teacher key, attach/map key file or exclude; never “AI trả lời giúp”
Conflicting keys	Both values with exam/key source links	Select evidence or enter corrected key; decision records actor and affected revision
Ambiguous numbering	Candidate section/question matches	Choose mapping with both contexts visible
Default points	Proposed value and affected item count	Set/confirm an explicit score, reviewing total impact
Unsupported interaction	Preserved original content and incompatibility	Intentionally transform with preview/confirmation or exclude; type relabel alone is insufficient
Lost formatting/object	Original rendition and missing semantic element	Repair supported representation, replace asset or explicitly exclude dependent content
Missing audio	Affected group and asset requirement	Upload/select usable audio or intentionally change the assessment with a recorded decision
Unassigned content	Exact source block and nearby context	Attach to question/material/instructions, classify metadata or exclude with reason

Selecting a finding focuses the field and source evidence. Resolving it updates counts but does not yank the teacher to a different question; “Tiếp theo” advances when ready. Editing an answer-relevant field invalidates affected confirmations.

Batch confirm exists only for equivalent low-risk findings, e.g. applying the same points to a visible listed set. The confirmation lists exact items, decision and total change. Blockers, uncertain answer mappings and unsupported content have no blanket dismissal. Filtered-out items are never silently included.

Keep text local while typing, serialize autosaves and show “Đang lưu” only when a request is in flight. “Đã lưu” means server acknowledgment for that revision. A save failure keeps the editor and unsent edits, shows retry and disables commit. Route/question changes await a flush or remain on screen with an actionable error.

A stale revision shows “Bản nhập đã thay đổi ở nơi khác” with compare/recover, never a blind reload. Saved revisions survive tab closure; unsent-outbox recovery must explicitly show its local status and ask which revision to apply on conflict. No correctness claim depends on beforeunload completing an async write.

New source/model/profile processing creates a candidate beside the teacher's existing draft. Default to preserving manual changes. Compare added/removed/moved questions, changed keys/content/material links and source coverage. Every applied change is scoped to a current draft revision. New or conflicting evidence can invalidate an old confirmation but cannot silently replace teacher text.

Offer per-item apply, keep current and inspect source. Equivalent low-risk changes may be selected as a visible set. Applying reruns relevant validators. A cancelled or failed reprocess leaves the old draft usable.

Preview uses the production learner renderer, with desktop/phone viewing modes, group materials, accessible tables, cloze links, images and shared audio placement. Use a teacher-authorized learner-safe projection; do not merely hide admin keys with CSS. Preview does not consume real student playback allowances.

The summary lists included/excluded questions, sections/groups, points, answer completeness, assets, defaults accepted and unresolved decisions. Every unresolved count links back to its filtered review set. Commit is enabled only for the current saved and validated revision; explanatory text identifies the remaining work.

“Tạo bản nháp đề” creates a draft, not a publication. During commit, show a stable pending state; disable duplicate clicks but also rely on server idempotency. On lost response, recover by import identity instead of offering a fresh commit. Success names the new draft and offers “Mở trình soạn đề” plus history. Do not unexpectedly navigate away while the teacher is reading the final summary.

The normal builder gets the same content editor and section → group → question structure, inline material preview and explicit shared audio policy. Restoring old versions copies the complete graph; bank insertion cannot orphan passage children. The builder now offers whole-group bank insertion with a destination selector and an independent save-to-bank action that retains the current editor. Empty groups stay selectable. Group navigation collapses to a content selector when the editor's available width is narrow; material/member content keeps its usable editing width.

During an attempt, the current group's material remains available across member navigation. Desktop offers a readable material area alongside the question; phone places it above the question with an explicit expand/collapse control and remembered state. Do not force repeated reading or destroy an in-progress answer when changing the material view. Group audio stays scoped to that group, with counts stable across navigation/reload. Answer controls remain stationary.

Check long passages, wide tables, picture captions, underlined option fragments, ordered cloze gaps and results with hidden answers. Student presentation must not expose source filenames/key documents, provenance, teacher notes or answer markings.

Results keep the group's material above its first visible child when filters change; following a material gap restores all questions and focuses its target. Teacher paper review keeps context while navigating members. Grading by question shows that question's complete material without combining different students' listening counts. Transcript disclosure is explicit and keyboard-operable; learners receive only released transcripts. Review playback never changes an attempt's counters.

• At ≥1280px: two full comparison panes, optional outline; avoid fixed empty space.
• At 768–1279px: collapsed app navigation, switchable source/result views when two readable panes cannot fit; preserve active edit, selection and scroll anchors.
• Below 768px: follow the app's admin support policy with accessible status/history and clear desktop/tablet guidance; don't promise dense editing on a phone without testing it. The interaction mockup still reflows to 320px for conversation use.
• Learner rendering remains fully supported from 320px. Long labels wrap; tables scroll within their own region without causing page-wide overflow.
• All actions work by keyboard. Native tab order, visible focus, labeled toolbar, movable dialogs, and no shortcut interception inside content editors. Prefer visible Previous/Next issue controls; add shortcuts only with discoverable help and tests against existing app bindings.
• Announce stage changes/save failures/validation results politely. Do not announce every keystroke or move focus after every background validation. Test actual screen-reader use and 200% zoom, not only automated accessibility scores.
• Use 150ms ease-out for pane/selection feedback and light shadows for raised controls. No page-wide entrance cascade, springing text, pulsing error badges or animation that moves an answer control. Honor reduced motion.
• Load source blocks/rendition pages by region and window large result lists; active editor stays mounted outside virtualization eviction. Keyboard navigation loads the target and restores focus. Mount one rich editor per active context.

Use authorized real documents for the pilot, synthetic documents for repeatable interaction tests. Record task success, active time, mistakes, recovery steps and teacher feedback. Do not claim a satisfaction score without collecting it.

1. Import a clean free-form exam without reading a template guide; reach learner preview and explain the difference between draft and published test.
2. Resolve contradictory same-file/companion answers with restarted numbering; point to both pieces of evidence and choose the intended question/key.
3. Fix a pronunciation underline and confirm it survives option editing and the phone learner preview without carrying teacher-only answer formatting.
4. Correct a shared reading/cloze group; move a child, undo/repair a broken gap, preserve the passage and explain which shuffle settings are compatible.
5. Attach one audio file to a group; see one allowance across its child questions.
6. Lose network during an edit, attempt to leave, retry, reload and recover; repeat with a second tab causing a revision conflict.
7. Reprocess a changed source after manual corrections; keep the corrections, inspect a conflict and selectively apply new evidence.
8. Exclude unsupported matching content with a reason; verify changed counts and totals, then restore it and see the blocker return.
9. Use only keyboard and a screen reader to navigate findings, correct a key and create a draft; repeat at 768px/200% zoom with Vietnamese and English labels.
10. Double-click commit or lose its response; recover the same draft without duplicates. Return from its builder and retain history/review filter state.

The first prototype demonstrates contradictory keys, score-default confirmation, source/result switching and the commit gate. Upload/reprocessing/audio/split-merge and genuine persistence require the later implementation and acceptance tests; do not infer their completion from the prototype.

The supplied corpus adds four concrete walkthrough cases: missing source underlines (offer correction without guessing), a 31-label exam with 35 answer targets (show the five-child group explicitly), a companion key containing several exams (confirm the matching exam and show unmatched entries), and an annotated homework containing learner answers plus teacher corrections (classify the evidence before accepting a key). Treat source defects separately from recognition defects in both wording and measurements. These are acceptance requirements, not functionality implemented by the initial source inspector.

R4 (T-R4.36 to T-R4.39 and T-R4.57) rebuilt the five import pages in TeacherLayout, on the deck in docs/design/deck/ and its tokens, with data-scale="deck" on each page root, in light and dark. Every string is a vi key first, then en. The behaviours below are read from web/src/app/router.tsx and web/src/features/imports/; the task notes in 74-r4.md have the measurements and the departures (DG-12 to DG-18, DG-21, DG-73, DG-142, DG-147, DG-148).

Route	Content width	Notes
/teacher/imports	1320	The history
/teacher/imports/new	860	Upload and paste
/teacher/imports/:id	720	Processing
/teacher/imports/:id/review	full, sidebar collapsed	Entering the review collapses the sidebar for the visit and leaving restores it (route handle sidebar: "collapsed"); it is no longer collapsed permanently
/teacher/imports/:id/confirm	1320	Preview and create

All five sit under ImportsGate, which answers from GET /teacher/imports/capabilities (§2). Another teacher's import answers as a missing one on every page. The breadcrumb names Tests, then Imports, then the page (the import's title on the review and the confirm page).

• Header. "Word and PDF files and pasted text you are turning into tests. Each import keeps its original content, its progress and your review, so you can come back at any time." Then the outline action "Paste a test" (/teacher/imports/new?source=paste) and the primary "Import from Word/PDF". Both show only while a new import can start, which is when intake and processing are both on; with processing off the page keeps its list and a calm note says why.
• Filters. One segmented control, a group of aria-pressed buttons and not a tablist (DG-17), in the order All, Processing, Ready for review, Couldn't process, Draft created, Cancelled, each with its count. The counts are the history's six facets, which follow the search and ignore the selected status. Processing selects awaiting_sources, queued, processing and committing. The search ("Search file or test name", 200px at least, 280px at most) matches the title and the current filenames, accents ignored; a pasted import is found by its title only. Status, search, page and size are in the URL, and a status change resets the page.
• Pager. The sizes are 20, 30 and 50, and real pagination shows whenever the list has a total.
• Grid and cards. The list is not a DataTable. A container query on the list itself (not the viewport) switches it: from 960px of its own width a six-column grid with table roles, minmax(200px,2.4fr) minmax(160px,1fr) minmax(150px,1.4fr) minmax(110px,1fr) minmax(150px,1fr) 32px, the deck's tracks with the fixed ones taken as minimums so longer Vietnamese labels fit (DG-21); below it each import is a stacked card with the tile, name, wrapping meta and the menu, then the status pill and the review chips, then the test link and the primary action. Nothing scrolls sideways at any width, and a row's controls exist once in the accessibility tree. At 1280px with the sidebar expanded the list is 976px wide, so a vertical scrollbar of more than 16px turns it into cards (DG-18).
• Cells. File: the tile, the name and "{creator} · {when} · with {key}". Status: the eight import statuses on the deck's five pills. To review: "{n} need action" and "{n} to confirm" from the review counts (an import whose draft was never assessed shows no counts). The test link. The primary action follows the status: continue the review, view progress, view the cause, open the test. The menu has Download original only.
• A pasted import shows the clipboard tile, the localized name "Pasted text" in place of the stored filename, and "{creator} · {when} · {n} characters".
• States. The skeleton, a filtered empty state with "Clear filters", the no-imports state with a hint that names both ways in, and a failed load with a retry that keeps the last list on screen. Active imports are polled only while the tab is visible and the user is active.

A "Source" control under the header, "Upload a file | Paste text", chooses the mode (?source=paste; anything else is the file mode). The options size to their content rather than the deck's fixed 168px (DG-21). One intake stays mounted across the switch, so chosen files, the pasted text and a typed title survive it.

• File mode. The Files card with Exam (required) and Answer key (optional), the deck's preflight messages and the server's limits, Test title prefilled from the exam file's name. "How to read the content" offers automatic recognition only. The privacy callout, "Where your content goes.", names files and pasted text and says only what the capabilities report.
• Paste mode. The page reads "Paste a test" and the "Test content" card takes the Files card's place, with no answer-key slot. The box is a plain monospace textarea at least 340px high. The counter reads "{n} of {max} characters", n counting code points of the NFC-normalised text (the text sent), max ImportLimits.pasteMaxCharacters; over the limit the counter and the border turn danger. "Paste from clipboard" replaces the text and moves focus to the box, and an empty or refused clipboard shows the deck's two toasts; "Clear" shows only with text; "Copy example" ships. The kind pill has one state, "Plain text · mark answers with * or "Answer:"", and nothing reads the clipboard's HTML (DG-12).
• Found so far. While the box has text: Sections, Questions, With an answer and Need an answer, then the notes (over the limit, no numbered question, which questions have no answer, a number seen twice in one part, images and audio are not pasted). The count is paste.ts, the server's grammar for text sources, run on a deferred value, never the deck's script (DG-13); api/testdata/pasted-text-counts.json holds it and the server to the same numbers.
• "How to lay out pasted text" is open while the box is empty: the deck's four tips beside the example (a container query on the card, 594px wide, DG-142), stacked below it otherwise. The example is a case of the shared fixture with nothing missing (DG-15).
• Title. The first non-blank line when it reads as one (at most 120 characters and not a part heading, a question or an option); a typed title wins; a blank title starts as "Pasted test · {date}" in the user's zone.
• Footer. Cancel and Start processing, which is disabled for an empty box, a text over the limit and a text with no numbered question, each with the deck's hint. When ready it reads "{n} questions found." and then the file flow's sentence, "Processing takes about a minute. You can leave this page.", until T-R4.51 measures the cold worker at 10 s or less for the deck's sample (DG-16). Ctrl or Cmd+Enter in the box starts, never during an IME composition.
• Start runs the file flow's steps with the text in place of the exam file: the capability re-check, createWordImport, pasteImportSource, processWordImport. A retry reuses the create and upload identities; an edited text takes a new uploadId; a 409 re-reads the import; a 429 and an intake refusal show the server's message.
• Leaving. Pasted text that has not started asks first (a router blocker and beforeunload); switching the Source does not ask, nor does the navigation after a start. The file mode has no guard: leaving mid-upload stops the intake quietly.

A 720px frame with the file tile (the clipboard tile for a paste), the name, "Started {when}" and "with {key}". The running card (role="status") draws the deck's five rows from the run's stage: Check the files, Read the content, Recognise the structure and Match the answers, Check the test. Elapsed time, "attempt n of m" and waiting-to-retry stay, with the callout "You can leave this page." Ready reads "We found {q} questions with {a} answers in {s} sections" with pills for "{n} need action", "{n} to confirm" and "{n} notes", each shown above zero, "Start review →" and Later. Failed names the stage, the error code in mono and the action: replace the exam file when the import accepts a new source, otherwise start a new import. Cancel asks first and states the configured retention; Cancelled is final; committing and committed are kept. A pasted import says "Check the text" and "{n} characters checked", then "Plain text · no formatting"; its failed state offers "Start a new import", which opens the paste mode; the cancel copy says "the pasted text". Its Ready card's first pill is "Every question has an answer" when nothing blocks, and the deck's "need an answer" pill is not built (DG-14). The processing history is absent until run events are exposed. The spinner pauses on hover and focus and is static under reduced motion.

One card the height of the viewport less 128px, at least 560px. The top bar has the title (an input, because it is part of the saved draft, DG-147), "{q} questions · {a} answers · {s} sections", the save state ("Saved", "Saving…", or "Couldn't save your last change" with Retry, role="status"), the severity pills and "Preview & finish →". The toolbar has the Section select, All / Needs action / To confirm / Info with counts, a Source | Test switch below 980px of content width, and the navigator ("{i} of {n} open", Previous, Next). From 980px the source pane and the test pane share a SplitPane (30 to 65 percent, 42 by default, arrow keys step 2, remembered under quizzivy.importReview.split). The source pane has Exam | Answer key, Extracted | Page view (Page view is off until a page renderer exists), Download original and the block tags. A pasted import has one option, "Pasted text", and one view, "As pasted". The test pane has the whole-test card, findings with their resolutions as radio cards, the Exclude dialog (a reason of up to 500 characters is required) and the Points dialog. The open question's prompt is the content editor of T-R4.63 in the import profile: "Insert gap" for a fill-in-the-blank, no Rich text | Markdown switch (import content is never Markdown), no Image or Audio. One card is open at a time, so one editor is mounted. A phone (below 768) is read-only with the deck's dismissible note. The stale reload and "Use the new result" notices, the leave guard and the lost-response commit copy stay.

/teacher/imports/:id/confirm: "Preview and create a draft" and "← Back to review". Two columns, minmax(0,1fr) 360px, from 1000px of the page's own width. The preview is the student's paper through StudentPreview with Computer | Phone and the answers hidden, and reads "Showing {k} of {n} questions…" when a type the paper cannot draw is left out (DG-148). The summary lists the answers included, sections and groups, answers set (with Fix), points (with Confirm), audio and images ("Images and audio are not imported. Add them in the test builder.") and what is still to resolve (with Review); only a row that blocks the draft is in danger ink. "Create draft test" is 44px and aria-disabled with its reason while off; a lost response replays the same request. The created draft's card offers "Open in test builder →" and "Back to imports".

The deck did not draw several behaviours the product has, and drew several controls the product has no backend for (DG-67). R4 keeps the first in the deck's styling and does not build the second. Their backend and UI are built in R6 beside T-R6.13.

Kept, in the deck's styling	Gated, absent until built in R6
Elapsed time, attempt n of m and waiting-to-retry on the processing page	The audio slot
The stale reload notice and "Use the new result"	Saved recognition profiles
The committed card and banner, and the lost-response commit copy	Source conventions
The leave guard on the review (and, for pasted text, on the paste box)	The template download
The no-imports empty state, with a hint that names both ways in	Move, Split and Merge in the review (Exclude and Restore stay)
Cancel as a final state; Download original in the row menu	The processing history; "Page view" in the source pane (shown, off)
Phones read-only, with the deck's dismissible note	"Create another test from this file", "Delete import" and "Process again"

"Paste a test" is not one of the gated controls: R4 builds it (§12.3) with its departures in DG-12 to DG-16. Its "How to read the content" card is the file flow's, so the same controls stay gated there.

---

_{Generated from docs/plan/18-word-import-ux.md. Edit that file in the repository; changes made on this page are overwritten.}