YZF Sudoku 使用手册 / User Manual

YZF Sudoku is a Web/WASM tool for logical solving, technique research, and training. It preserves the board, candidate state, conclusions, chains, highlights, and full solve path instead of merely returning a solution.

Major features include:

• Import of plain puzzles, Library strings, Sukaku, candidate grids, Sudoku Coach JSON, and SCv7_32 strings.
• Local browser OCR with side-by-side review and editable given/solved/candidate roles.
• Puzzle generation, technique-targeted training generation, and ER/EP/ED/YZF rating.
• Single hints, hint application, auto-solve paths, available-step search, path replacement, and dynamic explanations.
• Technique presets, individual toggles, Whip/gWhip memory modes, shortest-length comparison, and advanced sub-options.
• Manual coloring, circles, eliminations, chains, block marks, board capture, and mobile sharing.
• Batch generation, batch solving, and the TLG Truth-Link set-logic editor.
• A dedicated portrait/landscape mobile solve mode with touch-friendly marking.
• A Help & Workspaces center for autosave, OCR drafts, recent puzzles, task status, operation guides, language, appearance, and diagnostics.

Regular Web Build

Serve the regular build over HTTP/HTTPS. From the web-app directory, for example:

python -m http.server 8000

Then open http://127.0.0.1:8000/index.html. Double-clicking the regular index.html may cause the browser to block WASM, workers, or OCR assets.

Standalone Build

The single-file Standalone build can be opened directly. WASM, the OCR runtime, and models should already be embedded.

Main Interface

• Board: the 9×9 grid, givens, solved digits, candidates, chains, and highlights.
• Top entries: app status, Explain, Solve, and Help & Workspaces. Desktop may keep the language selector; narrow layouts move language, manuals, and other low-frequency entries into the hub.
• Main actions: Generate, Training Generate, Load, Undo, Redo, Auto Solve, Hint, Apply Hint, All Steps, and Fullscreen.
• Four tabs: Controls, Techniques, Solve Path, and Available Steps.
• Number pad: used for touch, pen, and Solve mode. It is hidden in the normal desktop layout after mouse input; OCR drafts may also show a Given/Solved role button.

Help & Workspaces

Use Help & Workspaces in the top bar as the unified low-frequency entry. It is a centered dialog on desktop, a bottom sheet in portrait mobile, and a right-side sheet in low-height landscape.

• Workspaces: restore the autosaved board and manual marks, an unfinished OCR review, or up to 12 recent puzzle snapshots. Recent cards include a 9×9 preview plus given, filled, candidate-cell, and manual-mark counts.
• OCR guide interaction: on the first correction step, tapping the highlighted cell proxies the tap to the real correction cell and advances automatically. The guide card’s Next button remains available as a fallback.
• Task Status: keeps generation, rating, auto solve, batch, OCR, saving, and PWA updates visible instead of reducing them to an unexplained spinner.
• Learn & Guides: open the bilingual Manual, Technique Guide, or replay Quick Start, Manual Marks, OCR Correction, and TLG Editing guides.
• Settings & Support: switch Chinese/English, choose Auto/Light/Dark/High Contrast appearance, and copy diagnostics that exclude puzzle contents.
• Language synchronization: Workspaces, Task Status, Diagnostics, guide controls, the Manual, and Technique Guide follow the main interface language instead of keeping separate language state.

Appearance and High Contrast

• Auto: follows the operating-system light/dark preference and updates immediately when it changes.
• Light: preserves the original bright analysis workspace.
• Dark: uses a dark analysis canvas for the board, candidates, forms, dialogs, Technique Configuration, Manual Marks, and TLG status while preserving the semantic colors of Truths, Links, and eliminations.
• High Contrast: emphasizes black/white boundaries, stronger focus rings, and shape differences while reducing shadows and decoration; it is not a simple color inversion.

The theme does not alter board truth, candidate coordinates, technique search, or TLG projection. Reduced-motion and forced-colors system settings are respected.

Desktop vs Mobile

Supported Inputs

• 81-character puzzle: 1–9 are givens; 0 or . is empty.
• Library colon format: preserves current candidate eliminations and user digits.
• Sukaku / candidate grid: a 729-character Sukaku string or 81 candidate-cell tokens.
• Sudoku Coach: JSON or an SCv7_32_ compressed string with givens, user digits, and candidate masks.

Loading

1. Open “More: Puzzle Input / Export / Rating”.
2. Put the input in the text area and press Load.
3. When Prefer Clipboard is checked, Load first tries clipboard text and falls back to the text area if reading or import fails.

If desktop clipboard permission is denied, disable Prefer Clipboard and load the text area directly. Puzzle Input inside mobile solve mode returns to the analysis view and opens the input panel.

Manual Entry from a Full Candidate Grid

1. Open the input panel and press Manual Entry. The editor starts with candidates 1–9 in all 81 cells.
2. Desktop: the role buttons primarily choose Given or Solved digit. Left-click a candidate to enter it, right-click a candidate to remove or restore it, use arrow keys to move, number keys to enter values, Ctrl/Cmd+number to toggle candidates, and 0/Delete/Backspace to clear a value.
3. Phone and tablet: the large Given, Solved digit, and Candidates buttons remain available. Select a cell, then use the large number pad; Candidate mode keeps or removes pencil marks.
4. After setting a given or solved digit, peer candidates update immediately. Clearing that value restores candidates that were only temporarily blocked, while preserving explicit manual eliminations.
5. Press Confirm & import. Row/column/box conflicts and empty candidate cells are checked before the state is imported as Sukaku or Library.

Manual Entry reuses the three-role OCR editor but does not save over or clear an OCR correction draft. Image OCR keeps its source-image comparison and recognized masks; it does not use Manual Entry's automatic candidate normalization.

Automatic Session Restore

The browser locally saves the current board, candidate state, language, technique configuration, and complete Manual Marks state. Session format v2 includes cell and candidate colors, circles, pending and applied eliminations, chain lines, construction lines, Mini-Regions, blocks, and the current marking tools, while remaining compatible with legacy v1 sessions.

Help & Workspaces can also resume an unfinished OCR correction draft and up to 12 recent puzzle snapshots. Recent puzzles are deduplicated, and the initial empty board is never added to the list. Clear Saved Session removes the board, language, technique data, and saved manual marks used for the next restore, but does not clear the board currently on screen.

Coach Candidate Masks

userCellCandidates contains 81 decimal masks separated by -. Digit d corresponds to 1 << d; candidates 2, 3, and 6 therefore use 4 + 8 + 64 = 76.

OCR runs entirely in the browser: images are not uploaded and no online fallback is used. Models load on demand; the first run may download sizable runtime/model assets, while later runs use local cache. Interrupted downloads attempt to resume when supported.

OCR Entries

• Select Image: choose a file on desktop or a photo from the mobile library.
• Take Photo: requests the rear camera; some browsers show a camera/gallery chooser.
• Read Clipboard: reads an image from the desktop clipboard.
• Ctrl+V: paste a screenshot when focus is not inside an editable field.

Replay OCR guide: replaying from Help & workspaces first attempts to read a clipboard image. On success it recognizes the image and enters correction directly. Otherwise it expands the existing image-input section and provides a top-layer Choose Image action. The formal OCR guide continues automatically after recognition; users do not need to enter correction first.

OCR Review Window

• The recognized image is shown on the left and a fixed 9×9 correction grid in the middle. Clicking either side selects the same cell.
• Switch among Given, Solved Digit, and Candidates. Givens are black large digits; solved digits are blue.
• In Given/Solved mode, a number replaces the cell; pressing the same number clears it. Candidate mode toggles pencil marks.
• Previous/Next moves selection; Clear removes the cell; Undo/Redo edits correction history; Reset OCR Result restores the original model output.
• Confirm & Import imports an editable OCR draft; Cancel discards the recognition.

Desktop and Mobile

Desktop and roomy landscape layouts use three columns and provide Fullscreen Review. Narrow phones stack the image, board, and controls vertically; the fullscreen button may be hidden. Opening OCR review exits mobile solve mode so landscape fullscreen remains in the correction editor.

Desktop keyboard controls include 1–9, 0/Delete/Backspace, arrow keys, Ctrl/Cmd+Z, Ctrl/Cmd+Shift+Z, and Esc.

Further Correction After Import

After confirmation, continue editing on the main board if necessary. An OCR draft adds a Given/Solved role button to the number pad so black givens and blue solved digits can still be corrected.

Recommended End-to-End Workflow

1. Load a puzzle: paste a string, use OCR, restore a session, or generate one. Verify all black givens first; OCR puzzles may need given/user-digit role correction.
2. Mouse input: work directly on the board. Left-click an active candidate to set it; left-click a user value to clear it; right-click a candidate slot to remove or restore it.
3. Keyboard input: use an arrow key to establish and move the selected cell. Press 1–9 to set a value, 0/Delete/Backspace to clear a user value, Ctrl/Cmd+1–9 to toggle candidates, Ctrl/Cmd+Z to undo, and Ctrl+Y or Cmd+Shift+Z to redo. Mouse input remains direct and depends on the clicked candidate position plus the left/right button; neither mouse nor keyboard reads the persistent on-screen digit selection.
4. Touch/pen input: the solve layout shows a single-row 1–9 digit strip. The next row provides Value/Candidate, Undo, Redo, Hint, Apply, Marks, and More; Clear is under More. Choose Value/Candidate and 1–9, then tap the target cell.
5. Analyze manually: open Manual Marks for circles, colors, chains, blocks, or staged eliminations. Visual marks are separate from the real candidate grid.
6. Check logic: Hint only previews the next program step. Press Apply Hint to commit it. Available Steps can compare multiple legal options.
7. Continue or recover: formal values, candidate edits, and applied hints enter Undo history. Undo mistakes before continuing.
8. Save the state: use Library, Coach, or Sukaku for the real candidate state; use Capture for colors, circles, chains, and blocks.

Value Mode

• Mouse: left-click an active candidate to enter it as a user value.
• Keyboard: use the arrow keys to select a cell, press 1–9 to enter a value, and press 0/Delete/Backspace to clear a user value.
• Touch/pen: choose a digit, then tap an empty cell.
• Entering the cell's current user value again clears it. Original givens are protected from ordinary solving edits.
• Candidates update after a value is entered. A rule conflict is reported as an invalid position.
• When a user value is cleared or replaced, the cleared cell is rebuilt from the imported initial candidate baseline filtered by current row/column/box legality. Only the removed digit is restored in related peers. Unrelated manual candidate eliminations remain untouched, and no candidate outside the original puzzle/Sukaku baseline is invented.

Candidate Mode

• Mouse: right-click a candidate slot to remove or legally restore it.
• Keyboard: Ctrl/Cmd+1–9 toggles the matching candidate in the selected cell without reading the on-screen Value/Candidate mode. Some browsers reserve main-keyboard Ctrl+digit; use the numeric keypad or the installed PWA/Standalone when necessary.
• Touch/pen: choose Candidate mode and a digit, then tap an empty cell.
• Candidate mode changes the real current candidate grid, affecting hints, auto solving, rating, and TLG.
• Hide Candidates changes display only; it does not delete candidates.
• Provisional/Formal Eliminations in Manual Marks are different. Provisional is visual only; Formal is committed only by Apply All Eliminations.

Hint, Apply, and Available Steps

• Hint: displays the technique, conclusion, structure, and highlights without changing the board.
• Apply Hint: commits the current placement/elimination and records it in Undo history.
• Available Steps: compares legal options for the same position. A normal click/tap previews only; path replacement is separate from editing the live board.

OCR Draft Rule

An OCR result may enter as an editable draft even before the givens form a unique puzzle. Correct it before using techniques that require uniqueness, especially UR, AUR, DAUR, and GUR.

Desktop vs Mobile

• Mouse: FB-style direct input is used and the number pad is hidden. Left-click sets an active candidate, left-click clears a user value, and right-click removes/restores a candidate.
• Mobile: ordinary solving should use the large “choose digit, then cell” flow. For precise candidate marks, select a cell and use the large digit buttons.
• When TLG editing is enabled, TLG owns board input. Disable it to resume ordinary Value/Candidate editing.

Normal Generation

Select an ER difficulty band, then press Generate. Available bands are Random, Easy, Medium, Hard, Unfair, Extreme, and Insane. The status reports clue count and rating.

Training Generation and OTP

The four generation controls stay on one line: Difficulty, Training technique, OTP, and Text filter. Difficulty uses a compact content-sized field so the technique selector retains most of the available width.

• OTP off: select a training technique and press Training Generate. The program generates and solves until the path contains that technique. The training solve strictly inherits the current technique switches and never enables unrelated techniques. If the target itself is disabled, the program first asks whether to enable only that target.
• OTP on with a selected technique: Basic techniques run to a stall, then only the selected technique is searched. The puzzle is accepted when one matching step leads to Full House, Hidden Single, and Naked Single all the way to the end.
• OTP on with no selected technique: after the Basic stall, all non-Basic, non-BruteForce logical techniques are searched; any matching STTE key step accepts the puzzle.

The OTP Basic set is Full House, Hidden Single, Naked Single, Locked Candidates, plus Naked/Hidden Pair, Triple, and Quad. On success, the live board opens at the Basic-stalled position for immediate key-step training.

OTP requires at least one valid key step, not a unique explanation. Multiple techniques, constructions, or chains explaining the same elimination do not disqualify the puzzle. Text filtering applies to the key step.

Rate Current Puzzle

Rate Current Puzzle always rates the board currently visible to the user, including entered values and the current candidate state. The initial puzzle type only chooses the execution thread: standard Sudoku is rated directly on the main thread for faster response, while a true Sukaku uses a Worker to avoid blocking the UI. Results normally include ER, EP, ED, and may include YZFRate.

• Hint: finds the next step under the current technique order, showing conclusions, highlights, and chains without changing the board.
• Apply Hint: commits the current placement/elimination and adds it to undo history.
• Explain: when the current hint or selected path step is explainable, the top Explain button opens a dynamic “why this works” tutorial.
• Auto Solve: computes a full logical path from the current board and opens the Solve Path tab. It is primarily a path viewer and does not permanently apply every step to the editable board at once.

Long tasks use a Worker in the regular Web build when available. Standalone or unsupported browsers fall back to the main thread.

Status may report solved, NoResult, Invalid, or a step limit. If an advanced technique is missing, verify that it is enabled and that the puzzle/candidate state is valid.

Solve Path

• Click a step to preview the board immediately before it and display its highlights. Expand nodes to inspect the description, nodes, edges, rank, and hashes.
• Multi-branch hints such as Force Chains expose branch controls. Desktop also supports ←/→ or [/].
• After path replacement, a panel shows old/new step counts, rating deltas, and an Undo Branch action.

All / Available Steps

All Steps collects multiple legal options for the current board. You can:

• filter by technique name, conclusion, coordinate, or description;
• filter by technique; the top of the technique list includes STTE (Single to the End), which shows only options after which Full House, Hidden Single, and Naked Single alone solve to the end. This includes only the first stage of FB checkSTTE, not lclSTE, BTE, or shorter-path options;
• use default order or sort placements/eliminations first;
• show only Replaceable options that can bind to the current solve path; when an explicit source-step number is absent, the board hash before the option is matched against the current path;
• press Clear to reset the text, technique, sort, and Replaceable filters together.

Preview and Replace

A normal click/tap only previews. For a Replaceable item, right-click on desktop or long-press for about 0.65 seconds on touch to replace the corresponding original step and recompute the remainder. An option identical to the current step is not replaced.

Preset Buttons

• All In: enables every implemented technique.
• High Speed: disables a group of the most expensive techniques while retaining broad coverage.
• Extreme Speed: disables more slow techniques and prioritizes speed.
• Whip Rating: basic techniques plus Whip/gWhip with the large-memory rating configuration. It returns the shortest Whip by default; the option below adds gWhip to the length comparison.
• Braid Rating: basic techniques plus Braid/gBraid.

Individual Settings

The technique table allows individual toggles and shows score/difficulty coloring. Complex AIC has a with AMSLS sub-option; JE has with JEPOM; MSLS has an Enable irregular MSLS search sub-option.

Whip/gWhip search memory modes:

• Auto: normal queues for solving, large memory for Whip Rating.
• Normal: prioritizes speed and memory use.
• Large: improves search coverage at higher memory cost.

Compare shortest length with gWhip: applies only when both Whip and gWhip are enabled. The solver finds the shortest Whip first, then searches only for a strictly shorter gWhip. Whip wins ties.

Technique settings are saved with the main session. Presets reset the Complex AIC/JE sub-options, so re-check them afterward.

Reading Highlights

Highlights may include placements, eliminations, structure cells, strong/weak links, ON/OFF nodes, ALS/AF regions, guardians, and branches. Open the top Technique Guide link for player-facing explanations in techniques.html.

Guardian Force Chain

Guardians of a dead pattern cannot all be false, so at least one is true. If every possible true guardian produces the same conclusion, that conclusion is valid. UR and Triplet Oddagon/Tridagon Guardian FC use the same forcing principle with different deadly patterns.

Open Manual Marks in the Controls tab. It is a human-analysis layer for hypotheses, teaching diagrams, and planned eliminations. Marks do not change the live board except through Apply All Eliminations.

Complete Workflow

1. Correct the puzzle and candidate grid, then select the area to analyze.
2. Open Manual Marks and choose a mode and color. Mouse input uses left/right clicks directly; touch and pen input use the paired action buttons in the panel.
3. Cell modes target the cell; candidate modes can target tiny candidates, or on mobile use a selected cell plus the large digit pad.
4. For a chain, select start and end. For a block, collect nodes and press Finish Block.
5. Use Provisional Eliminations while testing. Convert confirmed conclusions to Formal Eliminations, then press Apply All Eliminations.
6. Manual marks are saved with the main autosave session. Capture important diagrams for long-term archiving or sharing, because browser storage may still be cleared.

Mark Modes

Add, Erase, and Colors

• Add writes the current mark; Erase removes only the current mark type.
• Custom colors are stored in the current browser.
• Different mark types can coexist on one candidate. Reduce layers when the display becomes ambiguous.
• Chains, construction links, and Mini-Regions use paired actions: Strong/Weak, Construction Strong/Weak, and Green/Blue. Repeating the same type and endpoints removes that object.

Chains, Construction, and Mini-Regions

1. Mouse chains: choose the start candidate, then left-click endpoint two for strong or right-click it for weak. Construction mode uses the same buttons with orange links.
2. Touch chains: choose Strong/Weak or Construction Strong/Weak, then use “cell → large digit” for both endpoints.
3. Mini-Region: choose the start, then left-click endpoint two for green or right-click it for blue. Touch uses the Green/Blue buttons.
4. Remove or retype: same endpoints plus same type removes; same endpoints plus a different type retypes in place.
5. Reverse the arrow: the most recent start-to-end order becomes the displayed direction. Edge identity itself is undirected.
6. Undo the latest edge: Undo Line removes only the last edge in the chain list. Use step 2 for an earlier edge.
7. Cancel a mistaken start: press Cancel Start before choosing endpoint two. Changing mark mode also clears a pending start.

Mouse example

To remove a Construction Weak edge 2r1c1 → 7r4c5, choose Construction, left-click the start, then right-click the endpoint. Reversed endpoint order also works.

Mobile example

Choose the mode and paired type button, then tap the start cell and digit followed by the end cell and digit. Repeat the same type and endpoints to remove it.

Desktop

After mouse input is detected, the number pad, line-type selector, and Add/Erase buttons hide automatically. Touch or pen input on the board restores them, so a tablet with a mouse can switch between both rule sets immediately.

• Color Cells: left-click to add and right-click to remove.
• Candidate modes: left-click adds; right-click erases the current mark type.
• Candidate color/circle/eliminations: may target a 3×3 candidate slot even when that candidate has already been removed from the live grid.
• Chains/Construction: endpoint two uses left-click for strong and right-click for weak. Repeat the same type and endpoints to remove.
• Mini-Region: endpoint two uses left-click for green and right-click for blue. Repeat the same color and endpoints to remove.
• Blocks: right-click removes that candidate from the draft and every completed block containing it; empty blocks are discarded.

Mobile / Touch

• Color Cells: tap to add; hold the cell for about 0.6 seconds to erase.
• Candidate colors, circles, provisional/formal eliminations, and blocks: select a cell, tap a large digit to add, or hold it for about 0.5–0.6 seconds to erase. Direct long-press on a tiny candidate also works.
• Without long-press: choose Erase first, then tap the cell or large digit. Chains are the exception: remove a specific edge by repeating the same endpoints and type.
• Mobile chain/construction/Mini-Region: choose the paired type button, then use “tap cell → tap large digit” for both endpoints. Repeat the same type and endpoints to remove.
• Candidate modes automatically reveal candidates. Long-press erases only the manual mark, not the live candidate.
• In Solve Mode, press Marks. The panel stays inline in the solve layout instead of switching to a popup drawer. The main action row contains Value/Candidate, Undo, Redo, Hint, Apply, Marks, and More; Clear is available under More.

Operation Matrix

Chain and Block Tools

• Undo Line: removes the latest chain/construction edge, or the latest Mini-Region when that mode is selected. Redraw the same type and endpoints for a specific old object.
• Cancel Start: cancels the pulsing-yellow unfinished chain, construction, or Mini-Region start.
• Finish Block: commits the current candidate collection.
• Undo Block: removes the latest draft node first, or the latest completed block when no draft exists.
• Clear Marks: clears all manual colors, circles, eliminations, lines, and blocks without changing the puzzle.

Block Editing Boundaries

• The first node locks the current draft block color. Changing color before finishing does not recolor that draft; finish it and start a new block.
• One node renders a circle, two nodes a capsule, and three or more nodes a convex hull.
• Erasing one candidate removes it from the current draft and every completed block containing it, so overlapping blocks may all change.

Apply and Capture

Apply All Eliminations commits Formal Eliminations only, and only candidates that still exist at execution time. Stale marks on solved or already-removed candidates are skipped. Successfully applied formal marks leave the pending set; provisional marks remain.

With Clear Easy Steps enabled, easy-step cleanup is attempted even when there are no formal eliminations. It continues through Full House, Hidden/Naked Single, Locked Candidates, Pair, Triple, and Quad until a higher technique or no step is reached.

Capture creates a PNG exactly as displayed. Desktop first tries the image clipboard; mobile first tries system share; unsupported cases fall back to download.

Relationship with TLG

TLG editing and manual marks both need candidate touch ownership, so they cannot be edited simultaneously. TLG temporarily hides ordinary highlights, chains, and marks; disabling TLG restores them.

Batch Generation

1. Select Batch Generate, an output filename, and the main difficulty.
2. With OTP off, a selected Training technique emits ordinary training puzzles containing that technique. “Any AHS technique” accepts any currently enabled AHS-XZ, AHS-XY-Wing, or AHS-W-Wing target; if all three are off, the program asks whether to enable only those targets. No selection performs unrestricted generation.
3. With OTP on, batch generation reuses the main Training technique and text filter: a selected technique generates that technique's OTP puzzles, while no selection generates any-technique OTP puzzles.
4. Realtime normal training opens immediately before the target technique; OTP generation opens at the Basic-stalled key position. Temporary step highlights are not retained.
5. Press Start to generate continuously and Stop to finish and close/download the file.

Each puzzle occupies one tab-separated line with four columns: index, Library puzzle string, YZF score, and the SKFR ER score. When a training technique is selected, the batch panel defaults the filename to the technique name.

Batch Solving

1. Select Batch Solve.
2. Choose a plain-text input file with one puzzle per line. Blank lines and lines beginning with # are ignored.
3. Press Start. Output includes status, steps, YZF rating, hardest technique, and errors.
4. Hardest-technique tie-break: compare the YZF score first, then the fixed FB SolutionType order for equal scores; rank/length is compared only within the same technique. Thus, at score 400, Cell/Region Forcing Chain supersedes an earlier Complex AIC instead of depending on path order.

Files and Platform Differences

Desktop browsers with the File System Access API ask for a destination and append while running. Other browsers, commonly mobile and some Firefox/Safari versions, create a download when the task finishes or is stopped.

TLG Solver describes a set proof with Truths, Links, and optional uniqueness constraints, enumerates projections satisfying all premises, and derives common eliminations or assignments. It is not an ordinary hint button: the user defines the proof structure.

Enable TLG Editing gives TLG exclusive ownership of all board mouse, touch, long-press input, and its visual layer. Ordinary technique highlights, chains, and manual marks are hidden rather than deleted. Disabling TLG releases the temporary TLG candidate grid and computed result, while keeping the entered Truths, User Links, both Virtual Sets, AUR, DAUR, and GUR marks.

13.1 Four Different States

13.2 User Links, Automatic Links, and “Link Type = Auto”

• No User Links: Find derives automatic Links again from the current Truths, candidates, and extra constraints.
• Any User Link: Find uses explicit-link mode and does not mix in automatic Links.
• Result Links: the successful structure is stored separately for highlighting, regardless of whether it was automatic or explicit.
• Link Type = Auto: only controls how a pair of clicked endpoints is interpreted by the frontend. It does not enable backend-wide automatic Link generation.
• Successful Convert Redundant Truths or Remove Unused Links is an explicit structure mutation, so its returned Links become User Links. Remove all User Links to return to fresh automatic generation.

13.3 Candidate Grids and the Uniqueness Gate

TLG stores an initial candidate grid separately from the current candidate grid. The initial grid verifies AUR/DAUR/GUR premises; the current grid drives the proof. Sukaku is a format, not an automatic training-mode signal.

• Unique puzzle: verify initial swap pair (default): use for ordinary puzzles and unique Sukaku. AUR/DUR/GUR needs sufficient swappable/equivalent completions initially and a related live completion currently.
• Training grid: allow missing initial candidates: only for deliberately constructed multi-solution, Rank, or structure-training grids. It checks the current cloud without proving the original uniqueness premise.

13.4 Complete Workflow

1. Reach the candidate state to analyze and verify it. For a standalone candidate grid, paste 729-character Sukaku or 81 candidate-cell tokens.
2. Open TLG, choose the correct gate, and enable editing. Press Import TLG Candidate Grid when needed; imported candidates become the initial grid and later edits affect the current grid.
3. Enter Truths first. Use endpoint clicks for simple sets and the context menu for precise batch descriptors.
4. Choose the Link strategy: leave User Links empty for auto derivation, or enter explicit Links to reproduce a fixed structure.
5. Add Virtual Set, AUR, DAUR, or GUR only when the proof requires them.
6. Review Current TLG State for stale User Links, unrelated sets, or wrong candidates.
7. Press Find Eliminations and inspect Result Links, conclusions, Rank, expansion/validation data, and budgets.
8. Optionally run Convert Redundant Truths and Remove Unused Links. Failure preserves the previous valid state; success commits the returned structure.

13.5 Truth Input and Use

A Truth is a set from which exactly one member must be selected. Row-digit, column-digit, box-digit, and cell sets are supported.

• House Truth: in Truths mode click two same-digit candidates sharing the intended row, column, or box.
• Cell Truth: click the empty cell background in Truths mode, or choose Cell in the batch menu.
• Precise batch input: select candidates with Ctrl/Command or touch long-press, then choose Row, Column, Cell, or Box.
• Add only proven structural Truths. Extra Truths change Rank and the projection; more is not automatically safer.

13.6 Link Input and Use

A Link is covered at most once and also supports external-candidate elimination tests.

• Recommended default: leave User Links empty and let the engine derive them.
• Explicit Links: use for reproducing a specified TLG, constraining the structure, or manual Rank cleanup. House Links normally use same-digit candidates; a Cell Link uses candidates in one cell.
• Endpoint input: switch to Links, select Link Type, and click two candidates. Auto treats different digits in one cell as a Cell Link; otherwise the endpoints must use the same digit and share a row, column, or box. Endpoints only identify the descriptor and do not need to belong to a Truth; one endpoint may be outside the Truth footprint.
• Pending Links: manual construction is not gated by whether a Link produces an immediate conclusion. A Link with no current Truth intersection is retained so that Truths may be added later. Right-click or long-press one candidate to add a row, column, cell, or box descriptor directly. Only the explicit Remove Unused Links action prunes Links after proof-equivalence checking.
• Batch input is safer: explicitly choose Row/Column/Cell/Box when geometry could be ambiguous.
• One User Link is enough to disable automatic Link mixing for that Find. Delete all User Links to restore automatic mode.

13.7 Virtual Set

A Virtual Set expresses an exact cardinality that cannot conveniently be represented by one ordinary R/C/B/N Truth: exactly k selected candidates are true. TLG now supports two independent groups, Virtual Set 1 and Virtual Set 2, each with its own members and cardinality.

1. Select Virtual Set mode, then choose group 1 or 2 from Virtual Set Group.
2. Toggle member candidates for that group, or batch-toggle them through the right-click/long-press menu.
3. Set Truths to Apply to that group's exact cardinality k. Each group independently accepts 1–4.
4. Verify that every member is active and that external logic proves the exact cardinality of each active group, then run Find.

Relationship between groups: the two groups may be disjoint or may overlap. A selected candidate shared by both groups contributes to both cardinality equations.

Diagram: both groups use black set networks; small 1/2 tabs on candidate badges identify membership. A shared candidate displays both tabs.

Use it when: another argument proves exact 1–4 truths in one or two candidate clouds that are not ordinary house/cell sets.

Do not use it for: “at least one,” “at most two,” or speculative groups. Each group is a strict equality premise.

13.8 Fixed AUR

A fixed AUR supplies a proven swappable uniqueness structure. A standard four-cell AUR is compiled as maxBodyFootprint ≤ 3.

Standard Four-Cell AUR

1. Select AUR mode and AUR 1 or AUR 2. The two groups represent two independent structures.
2. For each group, mark exactly two diagonal candidate points with different digits, different rows, and different columns. They define the two rows, two columns, and body pair.
3. For example, 2r1c1 plus 7r4c5 defines rows r1/r4, columns c1/c5, and digits {2,7}.
4. Strict mode requires both alternating completions in the initial grid and at least one still live in the current grid.

Mark body-defining candidates, not guardians, target eliminations, or outside candidates.

Advanced generalized AUR: a group of 10–64 points enters generalized/rotating candidate-cloud validation. Use it only when the complete body cloud is known; ordinary UR/AUR work does not need it.

13.9 DAUR

DAUR is a dynamic AUR/DUR pool. The user supplies a small body-candidate pool; the engine expands every valid standard four-cell AUR and six-cell 2×3 / 3×2 DUR found inside it.

1. Select DAUR mode.
2. Mark active candidates that may belong to the intended AUR/DUR bodies. Exclude guardians, target eliminations, and unrelated candidates.
3. Start with the smallest clear pool. The result reports the number of expanded AUR/DUR constraints.
4. If no valid form expands, check missing body candidates, current live completions, and the initial swap premise under strict mode.

Use it when: several overlapping AUR/DUR forms may exist and manually specifying each one is inconvenient.

13.10 GUR: Purpose, Standard Workflow, and Verified Cases

GUR (Generic Uniqueness Region) validates template-free swappable deadly patterns. Within a user-defined local cell set, the engine enumerates legal one-digit-per-cell completions and compares their row-digit, column-digit, and box-digit exact-cover footprints. Two or more completions with the same footprint can be exchanged without changing the external constraints. GUR compiles those deadly completions into exact no-good constraints, then derives eliminations and assignments shared by every remaining full projection.

Typical uses: irregular deadly patterns, Qiu Deadly Patterns (QDP), some Reversed Deadly Patterns, and local uniqueness structures that cannot be represented accurately by fixed AUR or DAUR four/six-cell templates.

Standard Workflow

1. Start from an accurate candidate grid: use strict initial-swap validation for states derived from a unique puzzle. Use the training-grid gate only for deliberately constructed training candidates.
2. Define the local cell boundary: include only active cells that participate in the swappable completions. Exclude guardians, expected eliminations, and cells used only for external links.
3. Add one Cell Truth for every participating cell: a Cell Truth requires one selected candidate in that cell. Merely marking GUR candidates is not sufficient.
4. Select the complete GUR cloud: in a closed local model, batch-select any candidate in each participating cell and choose Toggle GUR Cells from the right-click/long-press menu to add every current candidate in those cells at once. When guardians or non-body candidates are present, use Toggle GUR Candidate Cloud afterward for candidate-level corrections. Missing candidates create false restrictions; unrelated candidates enlarge the completion space.
5. Keep User Links at zero for the first run: let the backend generate Links from the Truths, candidate grid, and GUR. Links displayed after the run are read-only Result Links and are not written back as User Links.
6. Run Find Eliminations: inspect GUR enumeration constraints, uniqueness-gate status, budgets, eliminations, and assignments. Enter explicit User Links only when reproducing or editing a specified structure.

Success Criteria

• At least one GUR enumeration constraint is reported, with no missing initial equivalent pair, no missing live deadly completion, and no budget overflow.
• The projection search completes. When a budget is exceeded, “zero eliminations” does not prove that no elimination exists.
• The intended elimination or assignment appears in the result. A colored GUR corner tab alone is only a membership marker, not a successful proof.
• A closed local model can produce additional eliminations or assignments beyond a tutorial's highlighted conclusion. This can be valid because TLG evaluates all complete projections, but the cloud boundary should still be audited.

Verified Case 1: QDP Type 1

This case uses 15 Cell Truths and 61 GUR candidates across 15 cells. The engine returns 26 Result Links, one GUR enumeration constraint, seven eliminations, and one assignment. The core QDP conclusion is r7c5<>26; the closed model further establishes r7c5=3.

:0000:x:.48.163.736...2.4.5..3.....+4.....+1.8...451+7..1.....+4..+7....9..4.2.1...736.478.21.:862 574::
15 Truths = {46N2 46N3 467N4 467N5 46N6 46N8 6N9}
26 Result Links = {2369r4 57b4 26c4 3789b5 236c5 5b6 235679r6 236b8 3r7}
GUR cloud = {
3r4c2 5r4c2 7r4c2 9r4c2 2r4c3 3r4c3 5r4c3 6r4c3 7r4c3 9r4c3
2r4c4 6r4c4 9r4c4 2r4c5 3r4c5 6r4c5 7r4c5 9r4c5 3r4c6 7r4c6
2r4c8 3r4c8 5r4c8 6r4c8 9r4c8 3r6c2 5r6c2 7r6c2 9r6c2 3r6c3
2r6c3 5r6c3 6r6c3 7r6c3 9r6c3 2r6c4 6r6c4 8r6c4 9r6c4 2r6c5
3r6c5 6r6c5 9r6c5 7r6c5 3r6c6 8r6c6 7r6c6 2r6c8 3r6c8 5r6c8
6r6c8 9r6c8 2r6c9 6r6c9 5r6c9 9r6c9 2r7c4 6r7c4 2r7c5 3r7c5 6r7c5}
GUR enumeration constraints = 1
7 eliminations = r7c5<>2, r4c5<>3, r6c5<>3, r7c2<>3,
                 r7c3<>3, r9c6<>3, r7c5<>6
1 assignment = r7c5=3

Reading the result: r7c5<>2 and r7c5<>6 are the QDP Type 1 target. The Cell Truth still requires one candidate in r7c5, so all remaining projections agree on r7c5=3. The other eliminations are additional common consequences of the same closed local model.

Verified Case 2: GUR Type 3

This case has 17 Truth descriptors expanding to 15 active cells, with 62 GUR candidates. The engine returns 31 Result Links, one GUR enumeration constraint, eight eliminations, and one assignment. The main conclusion is r8c3=1, corresponding to r8c3<>458.

:0000:x:1.2..+6348.4....5+76.6....+9122.6..4....3.6......1..5.+6......+61..9.+2.93.76.6....2..5:324 824 825 925 348 368::
17 Truths = {235678N1 2356789N3 23N6 1N4 1N5}
31 Result Links = {389r2 38b1 45789c1 45789c3 8b2 357r3 45789b4 134578b7 1r8}
GUR cloud = {
3r2c1 9r2c1 8r2c1 3r3c1 5r3c1 8r3c1 7r3c1 5r5c1 4r5c1 7r5c1
8r5c1 9r5c1 4r6c1 7r6c1 8r6c1 9r6c1 3r7c1 5r7c1 7r7c1 4r7c1
8r7c1 5r8c1 4r8c1 8r8c1 3r2c3 8r2c3 9r2c3 3r3c3 5r3c3 8r3c3
7r3c3 5r5c3 4r5c3 7r5c3 8r5c3 9r5c3 4r6c3 7r6c3 8r6c3 9r6c3
3r7c3 5r7c3 4r7c3 7r7c3 8r7c3 1r8c3 4r8c3 5r8c3 8r8c3 1r9c3
3r9c3 4r9c3 7r9c3 8r9c3 9r9c3 8r2c6 9r2c6 3r2c6 3r3c6 5r3c6
8r3c6 7r3c6}
GUR enumeration constraints = 1
8 eliminations = r8c9<>1, r9c3<>1, r8c3<>4, r3c4<>5,
                 r8c3<>5, r3c4<>7, r3c5<>7, r8c3<>8
1 assignment = r8c3=1

Reading the result: the three eliminations r8c3<>458, combined with the Cell Truth, establish r8c3=1. The other five eliminations are also common to every remaining legal projection.

Troubleshooting Order

1. Confirm that every participating cell has a Cell Truth and that the GUR cell set matches the Truth coverage.
2. Audit every cell for all allowed candidates, especially the target cell.
3. Delete all User Links and retry automatic Links, so a stale explicit structure cannot restrict the search incorrectly.
4. Check that the uniqueness gate matches the candidate-grid source and that the initial candidates were frozen before editing.
5. If a budget is exceeded, reduce the local region first; do not interpret the overflow as an invalid pattern.

13.11 Desktop Batch Menu

1. Hold Ctrl (Command on macOS) and click candidates. Click again to deselect.
2. Right-click an selected candidate.
3. Choose Row, Column, Cell, or Box under Truth/Link. If all generated descriptors already exist, the action removes them; otherwise it adds missing ones.
4. Virtual Set, current AUR group, DAUR, and GUR can also be toggled in batch. Toggle GUR Cells adds/removes all current candidates from the selected candidates' cells, while Toggle GUR Candidate Cloud keeps candidate-level precision.
5. Clear Candidate Selection clears selection only. Clear All Logic clears TLG logic/result but preserves the temporary TLG candidate grid.

13.12 Mobile / Touch TLG

• Short tap: executes the current mode. Truth/Link selects the first and second endpoints; VSet/AUR/DAUR/GUR toggles membership.
• Hold about 0.6 seconds: adds the candidate to batch selection and opens the TLG menu. Submenus open by tap rather than desktop hover; Toggle GUR Cells fills the selected cells with all current candidates in one action.
• Select several: long-press one candidate, tap outside to close, then long-press another. Long-press is additive; repeat and execute the batch action from the menu.
• Restart selection: choose Clear Candidate Selection. A normal short tap is not a batch-selection continuation; it runs the current mode and clears the previous batch.
• Cell Truth: in Truths mode tap empty cell background or an inactive candidate slot rather than an active candidate glyph.
• Accidental empty-slot input: except for that plain Cell Truth tap, right-click, long-press, Ctrl/Command-click, and clicks in other TLG modes are ignored and cannot alter the underlying Sudoku board.
• Manual-mark candidate holds are suspended while TLG is active, avoiding touch-handler conflicts.

13.13 Find, Convert, and Remove

• Find Eliminations: observational. It computes Result Links and conclusions without mutating User Links.
• Convert Redundant Truths: moves eligible Truths to Links on a valid result structure. Success updates Truths and User Links.
• Remove Unused Links: removes Links not needed by the proof. Success commits the simplified structure as User Links.
• Follow-up buttons require a valid elimination/assignment. A failed response preserves the previous valid structure.
• Any input edit invalidates old Result Links and conclusions, preventing stale highlights from explaining a new structure.

13.14 Highlight Legend

• Object colors: Cell light blue, Row purple, Column green, and Box brown; structure candidates use blue badges and final eliminations use red badges.
• Truth / Link: a Truth is a filled region and a Link is the hollow outline of the same geometry. Rows and Columns are continuous bands; a Box is a turning, branching orthogonal network.
• Cell: the region wraps the active candidate badges belonging to that Cell constraint rather than filling or outlining the whole cell.
• Candidates: a candidate shared by several Truths/Links still has one badge; the geometry underneath carries the constraint identities.
• Virtual Sets 1/2: black branching networks with small 1/2 membership tabs; AUR / DAUR / GUR: colored corner tabs rather than outer frames. Candidates shared by both Virtual Sets show both group tabs.
• Result priority: a final red elimination overrides structure blue; constraint regions are underneath and badges plus interaction accents are on top.

13.15 Result Notation and Budgets

Uppercase R/C/B/N denotes Truths; lowercase r/c/b/n denotes Links; N/n is a cell set. Status distinguishes User Links from Result Links and reports the gate, active candidates, AUR/DAUR/GUR validation, eliminations, and assignments.

13.16 Clearing Commands

• Delete an item in Current TLG State: removes that user input only.
• Clear All Logic in the context menu: clears all TLG logic and result while preserving the temporary candidate grid.
• Clear TLG State in the panel: also removes the temporary candidate grid; the main Sudoku puzzle remains.

13.17 TLG Logic Library

The TLG Library turns a candidate grid, Truth/Link structure, and uniqueness premises into a reusable study case. It is independent of the main-session autosave: the browser library is convenient for daily work, while an exported .tlgdb file is the durable backup for another device, reinstall, or long-term archive.

Fixed-Size Binary Records

The library uses a private YZF binary format with exactly 2048 bytes per record. After the file header, a record is located directly at “header + index × 2048,” making reads, in-place replacement, and ordered management straightforward. Every record has its own CRC32 checksum and the file header is checked as well. Damaged input is rejected before it can change the local library.

Each record stores:

• original givens, current placed digits, the initial candidate grid, and the current active candidate grid;
• 324-bit Truths, 324-bit User Links, and computed Result Links;
• two AUR groups, DAUR, GUR, and two Virtual Sets, each candidate cloud stored as a 729-bit bitmap;
• the independent cardinality of both Virtual Sets, the selected Virtual Set group, uniqueness gate, input mode, and Link type;
• the latest eliminations and assignments, plus title, tags, source, and study notes.

Read, Insert, Replace, Append, and Delete

• Read: restores the candidate grid, all TLG inputs, options, and the result snapshot, then enables TLG editing. The library grid is an independent TLG view; disabling TLG reveals the previous main Sudoku session again.
• Insert: saves the current TLG before the selected record. With no selection, it behaves like Append.
• Replace: overwrites the selected record while preserving its position and original creation time.
• Append: saves the current TLG at the end with a new record ID.
• Delete: removes the selected record after confirmation.

Title, tags, source, and notes come from the editor on the right. Binary schema v2 remains fixed at 2048 bytes; after reserving the second Virtual Set bitmap, UTF-8 limits are 128 bytes for title, 96 for tags, 128 for source, and 507 for notes. Legacy v1 records with 512-byte notes remain importable. Text beyond the current limit is rejected explicitly.

Binary Library Import and Export

• Export Selected: creates a .tlgdb containing only the selected record.
• Export Library: exports every record in the current order.
• Append to End: imports file records as new records at the end.
• Insert at Current Position: inserts file records before the selected record in file order; without a selection, it appends.
• Replace Entire Library: clears the local browser library and loads all file records after a second confirmation.

Sharing One TLG Case as Text

The binary library is for durable storage and bulk management. To share one case in a forum, chat, issue, or notebook, use a .tlg.txt text case. Text actions use the current TLG workspace plus the title, tags, source, and notes currently shown in the library editor, so an unsaved case can be shared directly.

• Copy Case: copies the readable multiline YZF-TLG-CASE:2 format. VIRTUAL1[k] and VIRTUAL2[k] store the two groups. Legacy YZF-TLG-CASE:1 remains importable.
• Copy One Line: copies compact YZFTLG2, using W1=/W2= for the two groups. Legacy YZFTLG1 remains importable.
• Paste Case: attempts to read the clipboard and opens the text box. If browser permission blocks clipboard access, paste into the box manually.
• Import Text: reads a .tlg.txt or ordinary .txt file into the text box.
• Export Text: saves the current TLG workspace as title.tlg.txt.
• Load Case: first shows a parsed summary of title, type, Truth/Link counts, active candidates, and conclusions. Only after confirmation does it replace the temporary TLG state. It does not save into the library automatically; use Insert, Replace, or Append afterward for durable storage.

A text case losslessly stores givens, current placed digits, both candidate bitmaps, Truths, Links, AUR 1/2, DAUR, GUR, Virtual Sets 1/2 with their independent cardinalities, Result Links, eliminations, assignments, options, and metadata. The multiline form uses readable candidate notation such as 7r3c4 and descriptor notation such as 1r3. The complete candidate grids use Base64URL bitmaps. Exported text includes CRC32, so truncated or unintentionally modified content is rejected.

Insane/superhard generation can draw randomly from the compressed built-in bank and import the decoded puzzle. The packed bank is much smaller than storing raw 81-character strings and is suitable for WASM/Standalone builds.

• Raw puzzle string: about 81 bytes per puzzle.
• Puzzle19 raw: about 19 bytes per puzzle.
• Compressed bank: actual size depends on the puzzle distribution and is normally smaller.

Puzzle Export Formats

• Original Puzzle: original givens only.
• Known Digits: all current large digits.
• Candidate Text: current candidate text by cell.
• Sukaku: the 729-character candidate grid.
• Library: board, solved digits, and candidate eliminations; best for reproducing the current state.
• To Coach: an SCv7_32 Coach compressed string.

Export Puzzle writes the result back to the input area and copies it to the text clipboard when allowed. The selected export format is stored locally.

Capture and Share

To preserve chains, circles, colors, and blocks, use Capture in Manual Marks rather than puzzle export. Desktop prioritizes PNG clipboard copy; mobile prioritizes system share.

At widths up to about 900px or with a coarse/touch pointer, the top bar shows Solve. It moves the real board and number pad into a fixed viewport rather than cloning them, so edits, Undo, hints, and analysis remain synchronized.

Recommended Mobile Solving Order

1. Load or generate a puzzle and verify the givens.
2. Select Value or Candidate, then choose 1–9 on the large pad.
3. Tap the target cell to enter a value or toggle a candidate. Tiny candidate glyphs do not need to be targeted for ordinary solving.
4. For human analysis, open Marks and use “select cell, then large digit” for circles, colors, or staged eliminations.
5. Use Hint for preview and Apply Hint only after confirming the conclusion.
6. Undo mistakes immediately; export candidate state or capture the diagram at useful milestones.

Top and Bottom Actions

• Back: first closes the current dialog, More drawer, or Manual Marks layer. It exits Solve Mode only when no inner layer remains. The badge on the title-bar Back icon shows the remaining depth.
• New Puzzle: chooses an ER band and generates. Existing progress/marks trigger a warning.
• Clear: clears the selected user value, or the selected candidate in Candidate mode.
• Undo/Redo: shares history with the main interface.
• Marks: opens the full Manual Marks panel.
• More: opens Hint, Apply, Available Steps, Puzzle Input, candidate visibility, same-digit highlighting, Screen Wake Lock, Fullscreen, Analysis Mode, Help & Workspaces, and language.

Title-Bar Status Icons

• Cloud: shows offline-resource download, full readiness, current offline state, update availability, or failure. Tap for details; when an update is ready, tapping switches to it.
• Disk: green means saved, amber means unsaved changes, spinning blue means saving, and red means failure. The session is flushed when the page goes to the background, closes, or is reclaimed by the OS.
• Task list: shows a count and active state while work is running; open it for per-task progress, elapsed time, completion, or failure details.
• Back arrow: explains what Android Back will close next. Its badge shows the current UI-layer depth.
• Temporary feedback: sharing, copying, session restore, and update completion appear briefly, then collapse after about 2–3 seconds.

Candidates and Manual Marks

• Ordinary candidate editing: choose Candidate and a digit, then tap a cell.
• Candidate marks: select a cell, tap a large digit to add, or hold it about 0.5–0.6 seconds to erase the mark.
• Cell coloring: tap to add and hold the cell for about 0.6 seconds to erase.
• Hide Candidates is display-only; candidate mark modes reveal them automatically.
• Long-press erase does not delete a live candidate. Formal eliminations still require Apply All Eliminations.

TLG on Mobile

TLG is an analysis feature. Open it from Analysis Mode and enable editing; candidate touch is then owned by TLG:

• Short tap: runs the current TLG mode. Truth/Link uses two endpoint taps; VSet/AUR/DAUR/GUR toggles membership.
• Hold about 0.6 seconds: adds the candidate to batch selection and opens the TLG menu. Submenus open by tap.
• Multi-select: long-press one candidate, tap outside to close, then long-press another. Selection accumulates until the batch action is executed.
• Clear batch selection: use Clear Candidate Selection. A normal short tap executes the current mode rather than continuing batch selection.
• Cell Truth: tap the empty cell background in Truths mode.

More Functions

• Hint / Apply Hint: Hint previews; Apply commits.
• Available Steps: leaves Solve layout and opens the Available Steps tab.
• Puzzle Input: leaves Solve Mode, opens the input panel, and focuses the text area.
• Hide/Show Candidates: changes display only.
• Disable/Enable Same-Digit Highlight: highlights large digits matching the current pad digit.
• Keep Screen Awake: enabled by default. Solve Mode requests a Screen Wake Lock and releases it when Solve Mode closes. Browsers release it in the background; the app requests it again after returning to the foreground. Turn the switch off to use the phone's normal sleep timeout.
• Fullscreen: calls the browser Fullscreen API.
• Analysis Mode: exits the dedicated solve interface.

Orientation, Images, and Cache

Landscape places board and controls side by side; portrait stacks and scales them. Normal fullscreen may enter Solve Mode automatically on small/touch devices; OCR correction fullscreen is exempt.

Mobile has no reliable Ctrl+V; use Select Image or Take Photo. Chrome Android has no exact Ctrl+F5. After installing an overlay package, close/reopen the page or clear the site cache if the old UI remains.

The regular Web build is now an installable PWA. It works on GitHub Pages, local HTTPS, and localhost; no custom domain or server backend is required.

Preparing the Complete Offline Package

1. Open the page online for the first time and wait for the cloud icon to change from blue download state to green.
2. Tap the cloud and confirm that it says “Offline resources are ready.” The roughly 20 MB package includes the main UI, Workers, WASM, OCR runtime, both OCR models, Techniques, and this Manual.
3. When Chrome exposes installation, the cloud says “tap to install the app.” Tap it and follow the system prompt to add the app to the Home screen.
4. Enable airplane mode, fully close the browser, and launch from the Home-screen icon. The cloud should report that all features remain available offline.

Cache and Update Rules

• A new Service Worker activates only after every resource has passed verification, preventing a new app.js from being mixed with old Workers or WASM. If preparation is interrupted, the complete previous release remains active.
• Incremental updates: files are compared by size and SHA-256. Unchanged files are copied from the previous offline cache without network access, so a small release normally downloads only changed JS, HTML, or Manual files.
• Resumable downloads: large .wasm and .ort files are checkpointed in 1 MiB chunks. Completed files and chunks survive an interrupted connection, closed page, or failed installation; reconnect and tap the cloud to continue from the first missing chunk.
• If the host does not support HTTP Range, the app automatically falls back to a whole-file download; unchanged-file reuse still applies.
• Shared ?p=s1.… URLs reuse the same cached app shell instead of creating one cached page per puzzle, and still restore the puzzle offline.
• The cloud turns purple only after the browser exposes a real waiting Service Worker. Tapping it opens a point-to-point MessageChannel handshake with that worker. If a staged file was evicted, the worker repairs the missing resource before activation. The page watches the handshake, worker state, and controllerchange, then reloads as soon as any authoritative signal confirms activation.
• Clearing browser cache requires preparing whichever resources were removed. The app detects missing files, turns the cloud yellow, and continues when tapped online. Clearing site data also removes update checkpoints, the saved session, settings, and the IndexedDB TLG user library.

Status Icons and Actions

The Standalone build runs offline without a local server. Its build script must embed or package the app JS, WASM, OCR JS, ONNX Runtime Web, runtime WASM, and both model files into one HTML file.

• The regular Web build can use Solver/Training/Batch Workers; Standalone falls back to the main thread when workers are unavailable.
• The browser may still restrict clipboard, camera, system share, and file pickers independently of offline operation.
• The single file is large, so first open and first OCR initialization may be slower.

If OCR works in the regular build but fails in Standalone, verify that ort.min.js, runtime WASM, puzzle_localizer.ort, and puzzle_classifier.ort are fully embedded, and that the OCR module has no remaining dynamic import() or local fetch().

Sudoku image recognition uses a local model trained by Alex Kubiesa / Sudoku OCR. Recognition is performed locally in the browser; images are not uploaded and no online fallback is used.

Recommended attribution:

Sudoku image recognition uses
a local model trained by
Alex Kubiesa / Sudoku OCR.
No online fallback is used.