All features and internal specifications of KataGo AI, plus UI/UX details, generated from Swift source code
Note: This page is a technical document generated by AI analyzing the app's Swift source code. Actual app specifications and behavior may differ in detail. Please be aware of this in advance.
Go Lab at a Glance: A One-Stop Experience from Play to Improvement
Go Lab is not just a Go app. The entire flow of "Play → End → Analyze → Review → Branch → Save" completes within a single app. No exporting to external tools, no switching to another app. The overview is shown below.
① Play→② End & Dead Stones→③ Analyze→④ Review & Branch→⑤ Save & Report
Phase
What You Can Do
Related Chapters
① Play
CPU / Local / Online. Supports HumanSL modern and classic styles, Lv.1–63, handicap stones, and playing as White.
How to read this page: Each chapter begins with "What You Can Do in This Chapter" (product introduction), followed by "Technical & Internal Specifications." Parts I–IV follow the product usage flow, and the final cross-cutting chapter covers design principles common to the whole app.
Part I: Play & Online Play
Chapter 1: Game Modes and Rule Settings
F-101 to F-107
What You Can Do in This Chapter: Play in three modes — CPU, local two-player, or online. Supports all combinations of Japanese, Chinese, and Korean rules with 9×9, 13×13, and 19×19 boards. Handicap stones and playing as White are also supported, and after the game ends, the AI automatically highlights dead stone candidates.
1.1 Game Modes
Mode
Overview
Network
CPU Game
Play against the KataGo engine. Choose difficulty, model, and side.
Not required
Local Two-Player
Two players alternate on a single device.
Not required
Online Play
Matched via the app's own server. Rating changes apply (Chapter 4).
Required
1.2 Rule Sets and Komi
Rules
Scoring
Komi
Notes
Japanese
Territory + Captures
6.5
Traditional Japanese method.
Chinese
Area (Stones + Territory)
7.5
Chinese method. Higher komi.
Korean
Territory + Captures
6.5
Similar to Japanese rules.
1.3 Board Size × Rule Combinations
Board Size
Japanese
Chinese
Korean
9×9
◯
◯
◯
13×13
◯
◯
◯
19×19
◯
◯
◯
All board sizes × all rules are supported.
1.4 Handicap Stones and Side Selection
Handicap Stones: Parses SGF AB/AW tags and places stones in the order "Black → White." When handicap stones are present, the side to move is set to White following SGF convention.
Side Selection: On the game settings screen, choose Black (first) / White (second). The AI automatically takes the opposite color.
Local Two-Player: No side selection is needed; Black moves first, then players alternate.
1.5 Game End and Dead Stone Judgment
Two consecutive passes or resignation ends the game. Under Japanese/Korean rules (territory + captures), a dead stone marking phase follows the game end.
Auto-Highlight AI Candidates: Immediately after the game ends, KataGo automatically highlights "dead stone" candidates on the board. Displayed with a semi-transparent red circle + red outline for clear visual indication.
Sensitivity Slider: A slider at the bottom of the screen lets you adjust the AI's judgment sensitivity (0.0–0.9, default 0.3). Intuitive control for "more candidates" or "fewer candidates."
Manual Correction: Tap highlighted stones to manually toggle life/death. Safe even when AI judgment doesn't match intuition.
Confirm: The "Confirm" button removes dead stones from the board and moves on to territory scoring. Removed stones are added to each player's captures.
1.6 Post-Game Actions
Result Modal: Displays the winner, territory breakdown (Black territory, White territory, captures, komi, total), and dead stone count.
Resignation Handling: When a game ends by resignation, it's shown as "Black/White wins by resignation" using dedicated wording. Territory breakdown is omitted, with only the winner and reason shown simply.
New Game Flow: The "New Game" button in the result modal starts the next game immediately.
— 01 —
Chapter 2: CPU Difficulty Control (Lv.1–63)
F-108
What You Can Do in This Chapter: Select CPU strength from 63 levels, Lv.1 to Lv.63. From beginner to pro-level, you can find an opponent at your level.
2.1 Difficulty and Internal Parameters
Difficulty is not a simple increase or decrease in Visits; it's implemented by adding intentional noise (randomness or mistakes) to KataGo's inference results.
Visits: Depth of search.
Random Rate: Probability of randomly choosing a move other than the best move.
Mistake Rate: Probability of playing an obvious blunder.
Is Weakening: Flag indicating whether weakening processing is applied.
2.2 Level Band Trends
Level Band
Approximate Strength
Internal Behavior
Lv.1–Lv.10
Beginner–Novice
High Random Rate and Mistake Rate; plays bad moves intentionally.
No weakening. Close to KataGo's pure inference results.
— 02 —
Chapter 3: HumanSL (Human-like AI)
F-109, F-110
What You Can Do in This Chapter: Use AI that chooses moves "likely to be played by a human of a specified rank" rather than "best moves." Choose between modern and classic styles, with separate strengths configurable for play and analysis.
3.1 Technical Characteristics of HumanSL
Single model, multi-rank support: Instead of preparing separate models for each rank band, a single model takes "rank" or "era" as input conditions, outputting a move probability distribution for any strength.
Profiles: Modern profiles rank_20k through rank_9d, plus classic profiles like preaz_5k.
Usable without search: The HumanSL network can select human-like moves with reasonable accuracy even without search (Visits), but reaching top human level requires search.
3.2 Two Profile Systems: Modern and Classic
System
Prefix
Characteristics
Best For
Modern Style
rank_20k – rank_9d
Learned from modern human players' records. Reproduces the rational move sense of the AI era.
Practicing with "human-like moves of today"
Classic Style
preaz_20k – preaz_9d
Learned from pre-AlphaGo human records. Greater freedom in the opening, with a traditional style.
Experiencing "human-like moves before AI," or enjoying the feel of classical Go literature
3.3 Separation of Play and Analysis Models
In Go Lab, you can set the play model and the analysis model separately, along with HumanSL profiles for each.
Play Model: The model used as the CPU opponent. When HumanSL is selected, specify a separate "play profile."
Analysis Model: The model used for review mode, position evaluation, and report generation. Higher-accuracy models (High / HumanSL) are recommended.
Example setup: "Play against HumanSL 15-kyu (human-like moves)" and "Analyze with High model (b18c384nbt) for accuracy" — this combination is possible. Play a game against HumanSL, then review it with a high-accuracy model — a complete learning cycle.
Note that when HumanSL is selected as the play model, b6c96 (Light) cannot be selected for analysis (the accuracy gap is too large). For analysis, choose from b18c384 or HumanSL.
Note
HumanSL doesn't aim for "best moves," so the position graph tends to be less stable than with standard KataGo. For high-precision analysis, use the High model (b18c384nbt).
— 03 —
Chapter 4: Online Play
F-111
What You Can Do in This Chapter: Play in real time against players worldwide via the app's own server. Ratings change based on results, and you can compete for the top of the leaderboard.
4.1 Matching and Ratings
Quick Match: Specify board size and rating band for automatic matching.
Rating: Glicko-2-based ratings. Updated appropriately based on wins/losses and opponent strength.
Nickname: Set freely before the game. Changes are immediately reflected on the server.
Leaderboard: Displays the top-ranked players. Season-based, with ratings reset each season.
4.2 Game Flow
Select "Online" on the game settings screen and start Quick Match.
On match, a 3.5-second countdown starts.
During the game, both players' moves are synchronized in real time.
On game end (two consecutive passes or resignation), the winner and rating change are finalized.
— 04 —
Chapter 5: OGS Integration
F-112, F-113, F-114
What You Can Do in This Chapter: Integrate with OGS (Online Go Server), one of the largest online Go servers. Both spectating (Watch) and playing (Play) are supported, and spectated games can be saved as SGF for analysis.
5.1 OGS Watch
Anonymous Authentication: View games and fetch lists without registering an account. Login-free spectating experience.
Live Game List: Fetches up to 100 games in progress, auto-updated every 30 seconds. Manual refresh button available.
Live Game Count Event: Subscribes to changes in the live game count and refetches the list only when it increases. When decreasing, relies on the 30-second timer, with an additional 10-second throttle to reduce server load.
Thumbnail Subscriptions: Only subscribes to game IDs currently visible on screen; unsubscribes when scrolled out. 0.5-second debounce batches updates, preventing excessive subscriptions during scroll.
Gamedata Watchdog: After starting to spectate, checks every 4 seconds whether gamedata has arrived. If it hasn't arrived for 10 seconds, auto-retries resubscription up to 10 times.
Clock Display: "Only the side to move has the elapsed time since server receipt subtracted." Uses half the RTT (up to 3 seconds) as a correction value.
AI Position Evaluation During Spectating: After the game ends, tapping the evaluation section evaluates each move with KataGo and accumulates the score array to update the graph in real time.
Save Game Record: Save spectated games as SGF for review mode analysis.
5.2 OGS Play
Quick Match: Specify conditions for automatic matching.
Open Challenge Creation: Specify your color (Black/White/Auto) and komi (Auto/Custom) to create a challenge.
Open Challenge Join: Tap another user's challenge to join. Wait for match notification, then go to the game screen.
Bot Games: Play against OGS bots. Bot-specific conditions are checked beforehand.
Rating: Glicko-2 rating changes.
Time Control Presets: Optimal presets are dynamically switched per board size.
Bot Compatibility Check: Verifies supported board sizes, time systems, speeds, and concurrent game counts. If concurrent games = 0, returns the error "This time setting is currently not accepted."
Rank Difference Limit: For ranked games, if the rank difference is 9 or more, the game cannot be accepted. Prevents excessively mismatched games.
Managing Your Challenges: Created open challenges are shown in the lobby as "Open Challenge" with a cancel button available anytime. Auto-cleared on match.
5.3 Time Control
System
Description
Fischer
Added increment per move on top of initial time (e.g., 5 min + 7 sec).
Byoyomi
After main time is used up, a fixed time is granted per move (e.g., 5 min + 5x30 sec).
Canadian
Complete a specified number of moves within a time (e.g., 5 min + 10 moves in 5 min).
5.4 In-Game Operations and Clock Correction
Illegal Move Blocking: Checks move legality on a virtual board reproducing the current position's rules; if it fails, does not send to the server.
Duplicate Send Prevention: Sets a "sending move" flag to block the next move's transmission. If the flag remains 10 seconds after sending, automatically resynchronizes the game connection.
Auto Timeout Send: When the clock event is received, if it's your turn and main time is -10 seconds or less, and byoyomi remaining is also 0 or less, sends a timeout once.
Detecting Opponent's Resignation: Handles game end when the opponent's last move is "resign" or on the game end event. Displays the result modal.
Rating Display: Fetches Glicko-2 rating and displays it alongside the rank, e.g., "5k (1500)".
5.5 Saving Spectated and Played Game Records
Unified History Saving: Both spectated and played games are saved locally in the same format. Can be reviewed later in review mode.
Title Prefix: Spectated games get the prefix "[OGS] " to distinguish them from regular games.
Metadata Preservation: Preserves both players' names, ranks, game name, result, komi, and rules.
"Save Only" Button: Always present at the bottom of the game screen. After game end, shows "Save Only"; after saving, shows "Saved ✓".
— 05 —
Part II: AI Analysis, Review & Game Records
Chapter 6: Real-time Analysis
F-201 to F-204, F-220-1
What You Can Do in This Chapter: View KataGo's analysis results in real time, during or after the game. Position graph, 5-tier candidate move color-coding, territory display, and thinking time highlights let you objectively assess the board state.
6.1 Position Graph
Vertical Axis: Score lead (in points). Auto-flipped so up is always favorable for you.
Horizontal Axis: Move number.
Current Position: Shown with a yellow marker.
Tap-to-Jump: Tapping the graph normalizes the tapped position to a move number and jumps to that position instantly (only in review mode).
Dynamic Y-Axis Scaling: Determines the upper limit by rounding up to the nearest 500 above max(2000, max absolute score + 200). Axis label step auto-switches between 5 / 10 / 20 points.
Dynamic X-Axis Step: Auto-selects steps that split the total moves into 7 parts (1, 2, 5, 10, 20, 25, 50, 100, 200, 500, 1000 moves).
Shared Y-Axis Between Main Line and Variations: The main line and variation tabs share the same Y-axis scale, so the visual height doesn't change when switching tabs.
Dashed Line for Branch Point: Draws a vertical white dashed line at the branch move number.
Previous Value Retention on Pass: When the score momentarily drops to zero (e.g., during komi calculation), retains the previous score to prevent graph crashes.
6.2 Candidate Moves (5-Tier Color Coding)
AI-recommended top candidates are displayed on the board.
Display Count: 3–30 moves (default 5). Adjustable in settings.
Color Coding: Judged by the worse of the win rate difference or point difference from the best move. Naturally color-coded with the same threshold in both the opening (points-dominant) and endgame (win rate-dominant).
Tier
Color
Condition (Difference from #1)
1
Blue
Best move
2
Green
≤ 2 points AND ≤ 2%
3
Yellow
≤ 5 points AND ≤ 5% (black text)
4
Orange
≤ 10 points AND ≤ 10%
5
Red
Beyond
Win Rate Digit Alignment: Separates integer and decimal parts; fixes the % sign to the lower right of the integer, stacking the decimal directly above. 99.2% and 99% don't shift in position and don't overflow the circle.
Full Zoom/Pan Tracking: Reflects the board's zoom and pan to compute each candidate's final position.
Proportional Shadow Scaling: Shadow offset is also calculated proportionally to the candidate's diameter.
6.3 Territory Overlay
Tapping the "map icon" displays AI-predicted territory on the board.
Setting
Behavior
Color only
Fixed size, expressed by intensity. Black territory opacity 0.05–0.65 / White 0.08–0.85.
Size only
Fixed intensity, expressed by size. Non-linear curve (exponent 0.65) emphasizes "weaker side differences."
Both
Expressed by both size and intensity.
Quantization: Does not draw territory with confidence below 0.3. Removes noise to show only "confident territory."
Outlines: Draws a thin white border for Black territory and a thin black border for White. Borders are visible on stones or wooden backgrounds.
Caching: Clears the previous map only on new game or board size change. No flicker even if temporarily empty between candidate evaluations.
Aggregate Values: Counts Black and White territory using confidence 0.6 as threshold, displaying as "●N ○M" next to the position graph header.
6.4 Thinking Time Visualization
Percentile Highlighting: In the game record list, sorts all moves' thinking times and colors the top 5% in red and the top 15% in yellow.
Live Update Granularity: HUD updates 20 times per second; cards at 0.4–0.5-second intervals. Balancing rendering load with visibility.
Accumulate Only During Turn: Adds only the current side-to-move's elapsed time. Values freeze after game end.
Recomputation on Replay: When jumping to a past move number, the HUD switches to "cumulative thinking time at that point."
— 06 —
Chapter 7: Review Mode
F-205 to F-208, F-220-2, F-220-3
What You Can Do in This Chapter: Tap the board from any position to try your own moves. Variations are managed by tabs, and PV ghost playback and variation trees give you a full overview. Jump to any position instantly from the position graph, game record list, or tree.
7.1 Creating Variations
Main Line Playback: Tap a move in the game record list to jump.
Create Variation: Tap the board to create a new variation and add a tab.
Switch Tabs: Switch between main line and variations using tabs at the top of the screen.
Delete Variation: Delete via the "×" button on the tab.
Variation Management Internals
Tab-Based Variation Management: Each variation holds "variation ID, all moves, current position, cursor position, and start move number" as a session.
Main Line Backup: On first variation, backs up the main line's all moves, position, and cursor position; accurately restores when returning to the main line tab.
Forward-Match Merge: When creating a variation, if the move sequence forward-matches an existing variation, moves to the existing tab without creating a new one. Prevents the accident of "duplicating the same sequence in a new tab."
Append vs. Mid-Branch: When the cursor is at the end, appends (extends main line/variation); when mid-sequence, creates a new variation.
Undo Mid-Branch: Deletes the end of the variation; when it drops to the variation start move or below, automatically deletes the variation tab and returns to the main line.
Safe Node Jump: Not a rewind but reapplies from the initial position up to the target move. Board state doesn't break even with complex variations.
7.2 PV (Recommended Sequence) Playback
PV Length: Choose from 3, 5, or 7 moves.
Playback: Displays PV moves in order on the board as semi-transparent + yellow-bordered ghost stones.
Place: Applies the PV moves to the actual board (confirms as a variation).
PV Preview Internals
Ghost Stone Visual Spec: Radius is cell size × 0.45, with shadow and a yellow 2.5pt border. White stones are opaque white circles; black stones are slightly transparent black. Move numbers drawn in the center in black/white text.
PV Operation Bar (Draggable): Combined into one floating bar: candidate move navigation, PV length switching, and Place button. Position is remembered across launches.
Bulk Apply: When Place is pressed, applies each PV move sequentially via the rules engine, generating new game record entries. Appends if at end, creates new session if mid-sequence.
Sequential Evaluation After Apply: Evaluates each move one at a time, filling in scores and updating the UI, so the position graph dots fill in one by one. If switched to another tab mid-way, evaluation stops immediately.
7.3 Variation Tree
Displays the full overview of all variations in tree format. Tapping any node jumps to that position. The rightmost column shows each move's "score difference" and "cumulative," letting you visually grasp position changes.
3-Column Right Panel: To the right of the board nodes, shows two columns: "Delta (position change by that move)" and "Total (position at that point)."
Bezier Branch Lines: Connects parent→child with a quadratic Bézier using the vertical midpoint as control point; active branches are drawn thicker (3.0pt vs 1.5pt).
Alternating Left/Right Layout: Variations alternate left/right, avoiding horizontal collisions even with deep branches.
Branch Tab Selection Scrolls to End: Tapping a variation tab scrolls to that variation's end (last move). "How far this variation has progressed" is visible at a glance.
7.4 Position Navigation (Cross-Jumping)
Go Lab provides multiple routes to access the same position from different screens.
Route
Starting Point
Destination
Position graph tap
Any position on graph
Jump to that move number
Game record list tap
Any move in record
Jump to that position
Variation tree node tap
Any node in tree
Jump to that position
7.5 Board Tap Control During Review Mode
Taps During Main Line Playback are Disabled with Hint: Tapping the board when review mode is OFF plays a warning haptic and shows a toast for 3.5 seconds: "Press the review button to place stones on the board and review." Prevents accidentally ruining the main line.
Taps During Review Mode: If the move progresses, a "click" mid-strength haptic; if not, a "buzzer" warning conveying "the move doesn't match existing moves / illegal move."
— 07 —
Chapter 8: Report Generation
F-209, F-210
What You Can Do in This Chapter: Analyze bad moves in two report types: "Score Drop" and "Winrate Drop." Includes comparison with AI-recommended moves, value gap display, and PDF/image save/share.
8.1 Two Report Types
Report Type
Target Move Extraction Conditions
Sort Key
Score Drop
Only moves by the specified side / Score drop ≥ 1.0 point / Does not match AI best move / Has evaluation
Descending score drop
Winrate Drop
Same as above + Winrate drop ≥ 2.0% AND score drop ≥ 1.0 point
Descending winrate drop
8.2 5-Tier Bad Move Evaluation
Evaluation
Score Drop Range
Meaning
Good
Less than 1.0 point
Close to the best move
Normal
1.0–2.5 points
Minor loss
Doubtful
2.5–5.0 points
Somewhat questionable
Bad
5.0–10.0 points
Bad move
Blunder
10.0+ points
Major blunder
8.3 Comparison with AI-Recommended Move
Best: Played move = AI best move (perfect)
Second: Evaluation difference under 2.5 points (sufficiently good move)
Mismatch: Move deviating from AI recommendation (room for improvement)
The report shows both the "played move" and "AI-recommended move" side by side on the board, and you can also check the PV (recommended sequence) for "what if this move had been played."
8.4 Value Gap
The position difference between the "played move" and "AI-recommended move" is calculated as a pure value gap compared at the same position. When the played move equals the recommended move, it's defined as 0.0 points.
8.5 Low-Confidence Flag
If the visit count when the move was analyzed is less than 500, a "Low Confidence" flag is added. This is a safeguard against declaring a move as "bad" when the analysis isn't sufficiently precise.
8.6 Async Analysis and Progress Display
Extract all candidate target moves
For each candidate, if the "AI best move" isn't cached, re-analyze with the engine, filling in best move, position value, recommended sequence, and visit count in the game record
Call a progress callback per move, showing "N / M" in the UI
Narrow valid candidates to the specified count, generating highlight info for each move
Calculate winrate change via sigmoid and embed
8.7 Report Output
Two Report Templates: Loads Score Drop / Winrate Drop HTML templates into WebView and renders JSON via JavaScript.
PDF→Image: Generates PDF in WebView, then rasterizes to high-resolution PNG. Height scales to max 4000px to save memory.
QR Code Embedding: Converts App Store URL to QR code, automatically inserting a link to "the app's store page" within the report image.
Clipboard/Share: Both clipboard copy and standard share sheet supported.
Copy Complete Toast: Shows a 1.5-second toast animation. Also plays a success haptic.
— 08 —
Chapter 9: Game Record Management & SGF I/O
F-211, F-212, F-220-2
What You Can Do in This Chapter: Game history saved as one JSON file per game. SGF import/export lets you import records from other apps for analysis. Autosave lets you resume from the last move even if the app crashes.
9.1 SGF Import/Export
Import (File Selection): Select an SGF file to analyze.
Import (Text Paste): Directly paste SGF text. Supports copy-paste from other apps.
Automatic Encoding Detection: Detects BOM and reads as UTF-8 first → if that fails, tries UTF-8 / Shift-JIS / UTF-16 in order.
Handicap (AB/AW) Support: Parses SGF handicap tags ([dp][pd]...) and processes per SGF convention: "apply Black→White in order" and "if handicap stones present, side to move is White."
Legal Move Check: Validates each move on import via the rules engine. If a move fails, logs and continues — a single failure doesn't lose the whole record.
Bulk Analysis After Import: Evaluates each move with KataGo, showing "N / M" via progress callback. Already-evaluated moves are skipped for resumability.
Export: Exports the game record in SGF format (GM[1]FF[4]). Can also be exported to other apps.
9.2 Game History Management
JSON File History: Saves as one JSON file per game in the app's documents folder. Second and subsequent loads are memory-referenced for speed.
History List: Displays date, player names, result, board size, and rules. Tap to go to analysis mode.
History Deletion: Swipe left to delete.
Title Prefix: Spectated games get the prefix "[OGS] " to distinguish from regular records.
Metadata Preservation: Preserves both players' names, ranks, game name, result, komi, and rules.
9.3 Autosave and Resume
Autosave: Saved every move during a game; a resume prompt appears on app restart (expiration: unlimited).
Resume Hint Elapsed Time Format:
Same day < 1 min: "Just now"
Same day < 60 min: "N min ago"
Same day ≥ 60 min: "N hr ago"
Past days: "N days ago" (calendar day difference prioritized)
Old Data Warning: When loading a record saved before the release date, shows a toast: "Data may be outdated." Saves up to the last 50 "toast-shown keys" so the same record doesn't show twice.
Migration from Old Data: Old-format history data is automatically migrated to file format. Deletes old keys on success, retries on next launch on failure.
9.4 Position Jump from Game Record
Game Record List Tap: Tap any move in the game record list to jump to that position (in review mode).
Variation Tab Switching: Switching between main line/variation tabs restores that session's cursor position.
— 09 —
Part III: Learning, Lessons & Tsumego
Chapter 10: Tsumego
F-301 to F-306
What You Can Do in This Chapter: A large collection of practical Tsumego problems. Filter by move count or status; the AI plays responses. Multiple solution visualization, solution playback, and streak features support daily learning habits.
10.1 Problem Structure
Collection: The currently active collection is based on "Xuanxuan Qijing (XXQJ)." Only problems with prepared SGF data are published; other classical collections will be added as data preparation is completed.
Filter: Filter by move count (1-3, 4-6, 7-10, 11+ moves) or status (unsolved, solved).
Sort: By problem number or by move count.
10.2 Solution Judgment and AI Response
Correct: Playing the correct move triggers a 0.5-second delay before the AI plays its response. This ensures the user's move has time to render, preventing inconsistencies from rapid taps.
Incorrect: Playing the wrong move shows "Failed" and the move is undone. The undo trigger is managed by a dedicated flag.
Free Play: For problems without a solution, play freely against AI (up to 15 moves). The AI is restricted from playing "outside the playable rectangle."
10.3 Multiple Solution Visualization
Multiple Solution Display: When there are multiple correct answers, playing one shows the other solution candidates as green dashed circles on the board. A highly educational effect that reveals "there were other options."
10.4 Solution Playback Mode
Solution Playback: After resetting the problem, lays out the solution moves one by one on the board. Four navigation options: step forward, step back, go to start, go to end.
Auto-Playback: Plays solution at 1-second intervals. Auto-stops at the end.
Move Number Overlay: Overlays "1, 2, 3…" on the stones. Black stones get white-background + black numbers; white stones get black-background + white numbers. Circle size is capped at stone diameter × 0.65 so the stone's edge is visible and Black/White can be distinguished.
Exact Board Coordinate Match: Shares the same coordinate calculation as board rendering, so move numbers follow the correct intersections even when zoomed/panned.
10.5 Progress Management and Streaks
Level-Specific Completion Data: Beginner/Intermediate/Advanced levels each store "solved problem numbers" separately.
Streak (Consecutive Learning): Stores "last practice date" and "consecutive days." Solve on the next day for +1; a gap of 1 day or more resets to 1. Multiple problems solved the same day don't increase the streak.
Today's Clear Count: Stores all "problem ID → solve datetime" and counts only those matching today's date.
Progress Notification: Sends a notification to related screens on clear, auto-redrawing the list.
10.6 Correct Answer Celebration
3-Stage Animation
Checkmark Spin-In: Rotates from -180° → 0° while springing from scale 0.2 → 1.0.
Star Burst: 14 stars scatter radially. Angle is 2π/14 with ±0.15rad jitter; distance 90–160pt; size 12–24pt.
Message Pop: "Solved!" message and buttons pop up with spring. Streak count shown as 🔥 N day streak in an orange capsule.
* Internal name: "Kururin~Pa! Kira~n" (spin-pop-sparkle). Presented as "Correct Answer Celebration" in the product introduction.
— 10 —
Chapter 11: Lessons
F-307 to F-310
What You Can Do in This Chapter: Learn Go from basics to advanced topics with a Live2D character as your guide. A "Read → Try → Solve" three-stage structure with board-interactive materials and tests enables active learning.
11.1 Lesson Curriculum Structure
Levels: 4 stages: Beginner / Intermediate / Advanced / Master. Only levels with prepared data are progressively published.
Lesson (Read): One lesson = multiple steps. Each step is linked to "board," "marks," and "captions," explained by a Live2D character.
Interactive Material (Try): Marks on the board are highlighted one at a time at 0.6-second intervals, visually following the order "look here → next here." Tapping during the typing animation skips to full display.
Test (Solve): An "interactive test" follows each lesson.
Tap All: Tap all correct points (with remaining count display).
Tap One: Tap one correct point.
On correct: light haptic + "Correct!" display. On incorrect: error haptic + "Try again" display. Incorrect messages are NOT auto-dismissed — they remain until tapped closed.
Completion Animation: Completing all questions shows a completion overlay, with "Next Lesson" to proceed.
Curriculum Intent: Rather than just "read and understand," interactive tests where you actually tap the board between lessons encourage active learning instead of passive reception. The "multiple solution dashed circles" mechanism is shared with Tsumego, showing other solution candidates when there are multiple correct answers.
11.2 Live2D Character Integration
Live2D Cubism SDK Integration: Dedicated wrapper bridges Live2D's Metal rendering View to SwiftUI. Singleton holds the ViewController, preventing flicker from multiple instantiations.
Lip Sync: Updates the character's mouth-movement parameter at 0.1-second intervals. Animates only for the first 3 seconds of dialogue, then closes the mouth.
Optimal Display Size: Live2D view height set to 1.5× the screen and extended to the bottom, so the caption background doesn't hide the character's feet.
Gaze Tracking: Leverages Live2D's own tap processing; SwiftUI only picks up black background areas. Designed so the overlaid transparent view doesn't block Live2D dragging.
Startup Warm-up: Executes model loading and Metal initialization 1.5 seconds after app launch. Eliminates the "character not displayed at the moment the lesson screen opens" wait time.
11.3 Board-Linked Explanations and Captions
Typing Animation: Displays captions with 40-millisecond-per-character typing. Includes a "skip" feature that instantly shows full text when tapped during typing.
Sequential Mark Highlighting: Highlights marks on the board one by one at 0.6-second intervals. Only active during lessons; disabled during tests, completion screens, and typing.
Highlight Visual Effect: The current mark scales to 1.3×; others dim to opacity 0.5. Smoothly transitions with 0.4-second easing.
Mark Color Coding:
Circle = Blue (point of interest/correct answer)
Triangle = Red (opponent's escape route/target point)
Square = Green (your shape/territory)
× = Orange (vital point/unplayable point)
Label = White
Inverted Stone Color: When a stone is at the mark position, the mark is drawn in the opposite color from the stone (White stone → black mark, Black stone → white mark).
Last Move Double Ring: Overlays "black thick outline (7pt) + white circle (5pt)" on the last move. Makes "where the move was just played" obvious during the lesson.
11.4 Level Management and Multilingual Support
4 Levels: Beginner / Intermediate / Advanced / Master. A mechanism limits currently published levels, hiding levels without prepared data from the UI.
Level-Specific Completion Data: Managed by separate completion data keys per level.
Multilingual Text: Supports Japanese / English / Korean / Traditional Chinese, selected at runtime based on device language settings. English fallback available.
Problem Number Interpretation: Extracts "problem (\d+)(?:-(\d+))?" from SGF comments via regex, decomposing into "(chapter, number)."
— 11 —
Part IV: Engine, System & Data Management
Chapter 12: System Overview & Requirements
F-401
What You'll Learn in This Chapter: An overview of how Go Lab is technically structured and what devices it runs on.
12.1 System Requirements
Item
Requirement / Specification
Supported OS
iOS 15.0+ / iPadOS 15.0+
Supported Devices
iPhone, iPad (Apple Silicon devices recommended)
Storage
~1GB+ free space when using High model
Network
Required for online play, OGS watch, and model download
Offline Operation
CPU games, Tsumego, and lessons work fully offline
12.2 App Structure
KataGo Engine: Calls C++ KataGo via Swift/Objective-C++ bridge. Runs inference on CoreML (ANE) or Metal (GPU) backend.
Game Controller: Manages game progress, Undo, autosave, end-of-game judgment.
UI Layer: Built with SwiftUI; board rendering uses Metal-based Canvas.
Persistence Layer: Saves game history and settings as JSON locally.
— 12 —
Chapter 13: AI Engine "KataGo" Technical Specifications
F-402, F-402-1, F-403, F-404
What You'll Learn in This Chapter: Technical specifications of the KataGo engine at the core of Go Lab: three models, backends, and inference parameters.
13.1 Model Types and Technical Differences
Model
Backend
Parameters
Features
Light (b6c96)
Metal (GPU)
~1.03M
Bundled with app. Lightweight and fast. g170 generation.
High (b18c384nbt)
CoreML (ANE)
~26.33M
Additional DL required. High inference accuracy. kata1 generation.
HumanSL
CoreML (ANE)
~26.33M + human imitation layer
Additional network trained on human game data, based on the High model.
* Parameter counts based on the publicly available KataGo network architecture.
13.2 Model Selection and Availability
Light: Bundled with the app, always available.
High / HumanSL: Additional download required. When not downloaded, the app automatically falls back to Light.
Selected Model: Switchable from the settings screen. On change, the shared engine is destroyed and the next inference uses the new model.
13.3 Backend Differences
Metal (GPU): Suitable for lightweight models like b6c96, running fast with low power consumption.
CoreML (ANE): Uses the Apple Neural Engine. Efficiently runs large models like b18c384, achieving high inference speed.
In simulator environments, Metal + b6c96 is automatically used.
13.4 Engine Inference Parameters
Visits during play: Dynamically set internally according to difficulty (Lv.1–63).
Visits during analysis: Adjustable from the settings screen in the range 40–1000 (defaults: evaluation 100, candidates 280).
— 13 —
Chapter 14: Settings & Data Persistence
F-405, F-406
What You Can Do in This Chapter: View the app's settings (Engine / Display) in one place. Learn about the data persistence method and autosave mechanism.
14.1 Settings List
Tab
Item
Range / Spec
Engine
Model in use
Light / High / HumanSL
Engine
Model download/delete
Additional DL/delete for High/HumanSL
Engine
Analysis Visits (evaluation)
40–1000 (default: 100)
Engine
Analysis Visits (candidates)
64–1000 (default: 280)
Engine
HumanSL profile
rank_20k–rank_9d / preaz_*
Display
Show lesson tab
ON / OFF
Display
Board style
Wood / Plain
Display
Territory display
Color only / Size only / Both
Display
Candidate move count
3–30
14.2 Data Persistence
Game History: Saved as one JSON file per game in the app's documents area.
Settings: Saved in the user settings area.
Autosave: Saved every move during a game; resume prompt on app restart (expiration: unlimited).
14.3 Autosave Robustness
Sanitize: Extreme expiration settings (like infinity) crash on JSON encoding, so they're replaced with a "effectively 100 years" constant for safe saving.
Broken File Evacuation: On decode failure, renames to .broken extension and evacuates. Other normal files continue loading.
Migration from Old Data: Automatically migrates old-format history data to file format.
Support: If you encounter an unresolved issue, please contact support ([email protected]). Include the app version, device name, and details of the problem.
— 14 —
Chapter 15: Model Management & App Operations
F-407, F-408, F-409
What You'll Learn in This Chapter: Mechanisms for downloading/deleting AI models, review request rules, and release note / forced update behavior.
15.1 Model Download Manager
Two-Stage Download: When HumanSL is selected, automatically downloads two files in order: base model (b18c384) → HumanSL model. Queues them for sequential processing.
Automatic .mlpackage → .mlmodelc Compilation: Compiles downloaded .mlpackage with Core ML and places it as .mlmodelc. Next launch loads the compiled model directly for faster startup.
Safety During Download: During download/delete, temporarily falls back to lightweight model (Metal b6c96). The app doesn't freeze even during heavy file processing.
Delete Dependencies: Deleting b18c384 also deletes the HumanSL model. Deleting only HumanSL keeps the base model.
Storage Breakdown Log: Logs app body, app data, cache sizes, and top 5 file breakdowns.
15.2 Review Request
3-Day Rule: Saves first launch date; suppresses review request if less than 72 hours since first launch. Consideration for users just starting.
Once per App: Manages "already requested" at app level. Whichever of game/Tsumego/analysis is triggered first, only the first passes.
Feature-Based Counting: Accumulates usage counts like "variation creation" and "SGF import analysis"; requests review on threshold.
Tsumego Milestones: Also requests review on reaching cumulative solves of 3 / 15 / 60 / 100 / 300.
15.3 Release Notes and Forced Update
Release Notes Fetch: Fetches App Store info, only queries when lastVersion != currentVersion. Adds simple language detection (Japanese: hiragana/katakana/kanji, Korean: hangul, Chinese: kanji) to skip "not the expected language" fallbacks. Prevents Japanese users from seeing English release notes due to fallback.
Forced Update: Numerically compares the minimum version on the server with the current version; shows an App Store prompt if outdated.
— 15 —
Chapter 16: Communication & Online Play Robustness
F-410, F-411
What You'll Learn in This Chapter: Mechanisms that keep games from breaking down even if the network becomes unstable during online play.
16.1 Background Departure Timer
When the app goes to the background, if in an online game, automatically leaves after 20 seconds. Timer cancels when the app returns. Consideration to not keep the opponent waiting.
16.2 Detecting Opponent Resignation and Game End
Handles game end when the opponent's last move is "resign" or on the game end event. Displays result modal with animation.
16.3 Move Count Sync
Compares server move count with local game record count, preventing "double application of already-applied moves."
16.4 Network Recovery
Pending Move Reconciliation: Compares "pending moves" with "server's last move"; resends if not yet reflected, clears if already reflected. Prevents both double sends and misses.
Disconnect Detection and Reconnection: Monitors network changes; on offline recovery or interface switch detected, waits 1.5 seconds before recovery. Design to avoid rapid reconnection storms.
16.5 Duplicate Send Prevention
Sets a "sending move" flag, blocking next move transmission. If the flag remains 10 seconds after sending, automatically resynchronizes the game connection.
What You'll Learn in This Chapter: Design refinements common to the whole app — board operation comfort, immersive effects, Zen Mode, physics animations, and more.
17.1 Board Operation Comfort and Error Prevention
CPU Move Auto-Follow: When the board is zoomed in and the CPU (opponent) plays off-screen or at the edge, the screen automatically pans to follow so the move is visible. However, it doesn't follow "your own moves," preventing motion sickness and mis-taps.
Auto Zoom-Out on Play: Tapping an intersection while zoomed in briefly zooms back to 1× before placing the stone. No stress of "playing while zoomed in and not seeing the whole board."
Double-Tap Magnification Toggle: Just double-tap to instantly switch between "1× display" and "zoomed to show 12 intersections per short edge."
Pan Clamping: When zoomed, limits the movable range so outside-the-board blank space isn't overexposed.
Multi-Device Layouts: Supports iPhone portrait/landscape, iPad, and fullscreen. In landscape, an information panel is placed to the right of the board, auto-switching to optimal layout.
17.2 Immersion-Enhancing Effects and Feedback
Zen Mode: Gently fades out all UI except the board (game record, candidate moves, position graph, control buttons, etc.), letting you fully focus on the board. Toggle anytime with the "Zen" button in the upper right.
Stone Placement Physics Animation: "easeOutCubic" motion as if the stone falls from above, with shadow rendering that changes in tandem (spreads thinly in the air, converges darkly on landing) computed in real time.
Haptic Feedback: "Click" impulse when placing a stone (CoreHaptics), "ta-ta-ta-ta" warning vibration when tapping an unplayable spot — haptic support for operation.
Thinking/Analyzing Dot Animation: Shows a flowing "…" animation above the board's top center, conveying AI thinking or analyzing without being overbearing.
Board Style Change: Toggle between "Wood" and "Plain" in settings. Plain gives a flat, modern look, suited for concentrating on analysis and review.
17.3 Analysis & Review Conveniences
Graph Tap-to-Jump: In review mode, tap any position on the position graph to jump instantly to that move number.
Auto-Flip to Player Perspective: When playing White (second player), the position graph's up/down and score's plus/minus are automatically flipped. Always "is it good or bad for me" is intuitively clear.
Candidate Move Legibility: Win rate (%) and score difference are shown inside each candidate move circle, with custom rendering ensuring number positions don't shift when circle size changes.
PV (Recommended Sequence) Ghost Display: Turning PV on in review mode shows the recommended sequence as semi-transparent virtual stones (ghosts) on the board. The "Place" button applies them to the actual board (confirming as a variation).
Variation Tree Visualization: Displays all branches as a tree. Each move's "score difference" and "position" appear on the right, showing at a glance where the position worsened.
Smart Variation Merge: In review mode, tapping the board to play a new move auto-creates a tab, but if it matches an existing variation's sequence, it merges (moves) into the existing tab without creating a new one.
17.4 Learning & Lesson Refinements
Live2D Character: A Live2D character appears as the guide in the lesson screen. Lip sync (mouth movement) and blinking synced to dialogue are implemented.
Sequential Highlighting: In lesson explanations, marks on the board (circles, triangles, squares, ×) highlight in order at fixed intervals (0.6 seconds), letting you visually follow the explanation order.
Multiple Solution Visualization (Tsumego): When a Tsumego problem has multiple correct answers, playing one shows the other candidates as "green dashed circles."
Move Number Overlay (Tsumego): In solution playback mode, move numbers "1, 2, 3..." are shown on the stones. Black stones get white numbers, white stones get black numbers, with size adjusted so the stone's edge is visible.
Correct Answer Celebration: When you solve a Tsumego problem, a unique animation (checkmark, message, star particles) celebrates you.
Streak (Consecutive Learning) Feature: Solving Tsumego daily records consecutive days, showing "🔥 N day streak" on clear.
17.5 Online & OGS Integration Conveniences
Thumbnail-Filled Watch List: The OGS Watch list shows actual board thumbnails. Move count and player ratings visible at a glance.
AI Position Evaluation During Watch: Even on the OGS watch screen, you can display a position graph using your own KataGo engine.
Rate Limiting: Prevents exceeding OGS API limits (requests per hour) via the in-app rate limiter.
Time Control Presets: OGS game time settings can be intuitively chosen from recommended presets per board size (9×9, 13×13, 19×19).
17.6 Zen Mode Operation Scheme
What Remains: Board, Undo/Pass/Resign buttons, Zen exit button (thinly visible).
What Disappears: Header, thinking time, game record, position graph, candidate moves, territory display.
Floating Menu: A "≡" button floats in the upper right; tap to open the normal menu (game record and settings accessible even in Zen).
Sequential Fade: Header → thinking time → information cards, each disappearing from outside inward with 0.08-second staggering. Not disappearing all at once; naturally guiding the gaze to the board.
Extended AI Move Interval: In Zen, the minimum interval before AI plays is extended from 1.0s → 2.0s, letting you look at the board carefully while waiting for AI's move.
First-Time Guidance: On first launch only, shows a 10-second toast when closing the game settings sheet.
17.7 Other Fine UX/Support Features
Autosave and Friendly Resume: Resume from the last move even if the app crashes. On resume, friendly messages like "N min ago," "N hr ago," "N days ago" appear.
Old Data Warning Toast: Automatically shows a toast when loading old game records saved before the release date, warning "Data may be outdated."
Report QR Code: Generated analysis reports automatically embed a QR code linking to the app's App Store page.
Multilingual Support: UI switches per device language: Japanese, English, Chinese (Simplified & Traditional), Korean, etc.
Forced Update / Release Notes: Fetches version info from server at launch, prompting App Store navigation or showing release notes as needed.
17.8 Other Hidden Considerations
First-Frame Animation Issue Avoidance: Delays animation start by one frame. Prevents the "animation doesn't play" bug from SwiftUI's initialization order.
Large Game Record Scroll Optimization: Uses lazy-rendered lists and auto-scroll, scrolling to the current move's row whenever the move number changes.
Perceived Speed During Replay: When "play from mid-way," kicks off inference first and waits for completion before starting playback. Prevents "juddering" from advancing before inference completes.
Undo Thinking Time Restoration: Subtracts the move's duration from accumulated time, clamped to 0 minimum. Repeated Undos don't make accumulated time negative, and Undo→Redo stays consistent.
Legal Move Zero Fallback: If AI returns an illegal move, enumerates legal moves, picks one at random, and passes if none exist. Prevents the game from stalling due to AI bugs.
Model Destruction on Settings Change: When the model in use changes, forcibly destroys the shared engine. New engine is generated with new settings, preventing old model mixing.
Simulator Model Auto-Switch: Simulator environments automatically force lightweight model (Metal b6c96).
— 17 —
Appendix: Feature Master Ledger
The master ledger of all features covered on this page. Used as the source of truth for future updates. Status legend: ✅ Feature = directly usable by users / 🔧 Internal = internal behavior / 📄 Doc = documentation item / 📂 Category = parent feature ID (bundling children) / ❓ Unverified = code unconfirmed.
Part I: Play & Online Play
ID
Feature Name
Ch.
Status
Parent
Related
F-101
Game Modes (CPU / Local / Online)
1
✅ Feature
—
F-111, F-113
F-102
Rule Sets
1
✅ Feature
—
F-107
F-103
Game End (Two Passes / Resign)
1
✅ Feature
—
F-104
F-104
Dead Stone Judgment
1
✅ Feature
—
F-103
F-105
Handicap Stones (AB/AW)
1
✅ Feature
—
F-106
F-106
Side Selection (Black/White)
1
✅ Feature
—
F-105
F-107
Board Size × Rule Combinations
1
✅ Feature
—
F-102
F-108
Difficulty Lv.1–63
2
✅ Feature
—
—
F-109
HumanSL (Modern/Classic)
3
✅ Feature
—
F-110
F-110
Play/Analysis Model Separation
3
✅ Feature
—
F-109
F-111
Online Play
4
✅ Feature
—
F-101
F-112
OGS Watch
5
✅ Feature
—
—
F-113
OGS Play
5
✅ Feature
—
F-101
F-114
OGS Time Control
5
✅ Feature
—
F-113
Part II: AI Analysis, Review & Game Records
ID
Feature Name
Ch.
Status
Parent
Related
F-201
Position Graph
6
✅ Feature
—
F-220-1
F-202
Candidate Moves (5-Tier Color)
6
✅ Feature
—
—
F-203
Territory Overlay
6
✅ Feature
—
—
F-204
Thinking Time Visualization
6
✅ Feature
—
—
F-205
Review Mode (Variation Creation)
7
✅ Feature
—
F-206, F-207, F-208
F-206
PV Ghost Playback / Bulk Apply
7
✅ Feature
—
F-205
F-207
Variation Tree
7
✅ Feature
—
F-220-3
F-208
Forward-Match Variation Merge
7
✅ Feature
—
F-205
F-209
Report Generation
8
✅ Feature
—
F-210
F-210
5-Tier Bad Move Eval / AI Move Comparison
8
✅ Feature
—
F-209
F-211
Game Record Management (JSON History)
9
✅ Feature
—
F-212
F-212
SGF Import/Export
9
✅ Feature
—
F-211
F-220
Position Navigation (Cross-Cutting)
6, 7, 9
📂 Category
—
—
F-220-1
Position Graph Tap-to-Jump
6
✅ Feature
F-220
F-201
F-220-2
Game Record List Tap-to-Jump
7, 9
✅ Feature
F-220
F-205, F-211
F-220-3
Tree Node Jump
7
✅ Feature
F-220
F-207
F-220-4
Variation List → Variation Navigation
TBD
❓ Unverified
F-220
—
F-220-5
Variation Drill → Position Navigation
TBD
❓ Unverified
F-220
—
Part III: Learning, Lessons & Tsumego
ID
Feature Name
Ch.
Status
Parent
Related
F-301
Tsumego Problem Structure
10
✅ Feature
—
—
F-302
Solution Judgment & AI Response
10
✅ Feature
—
F-303, F-304
F-303
Multiple Solution Display
10
✅ Feature
—
F-302
F-304
Solution Playback Mode
10
✅ Feature
—
F-302
F-305
Progress Management / Streaks
10
✅ Feature
—
—
F-306
Correct Answer Celebration
10
✅ Feature
—
F-302
F-307
Lessons (Live2D)
11
✅ Feature
—
F-308, F-309
F-308
Board-Linked Explanations
11
✅ Feature
—
F-307
F-309
Curriculum Structure
11
✅ Feature
—
F-307
F-310
Level Management / Multilingual
11
✅ Feature
—
F-307
Part IV: Engine, System & Data Management
ID
Feature Name
Ch.
Status
Parent
Related
F-401
System Overview & Requirements
12
📄 Doc
—
—
F-402
KataGo Model Types
13
📄 Doc
—
F-402-1
F-402-1
Model Selection / Availability Check
13
✅ Feature
F-402
F-110
F-403
Backends (Metal / CoreML)
13
📄 Doc
—
F-402
F-404
Inference Parameters (Visits)
13
📄 Doc
—
F-108
F-405
Settings List
14
✅ Feature
—
—
F-406
Autosave / Resume Hint
14
✅ Feature
—
F-211
F-407
Model Download Manager
15
🔧 Internal
—
F-402-1
F-408
Review Request
15
🔧 Internal
—
—
F-409
Release Notes / Forced Update
15
🔧 Internal
—
—
F-410
Network Recovery / Duplicate Send Prevention
16
🔧 Internal
—
F-111, F-113
F-411
Background Departure Timer
16
🔧 Internal
—
F-111
Cross-cutting: UI/UX
ID
Feature Name
Ch.
Status
Parent
Related
F-501
Board Operation (Follow/Zoom/Pan)
17
✅ Feature
—
—
F-502
Physics Animations
17
🔧 Internal
—
—
F-503
Zen Mode
17
✅ Feature
—
—
F-504
Haptic Feedback
17
🔧 Internal
—
—
F-505
Other Hidden Considerations
17
🔧 Internal
—
—
Unverified Items: F-220-4 (Variation List → Variation Navigation) and F-220-5 (Variation Drill → Position Navigation) are marked "❓ Unverified" because the corresponding UI could not be confirmed in the provided source code. Once the implementation is confirmed, the status and chapter will be updated.
Note: This page is a technical document generated by AI analyzing the app's Swift source code. Actual app specifications and behavior may differ in detail. Please be aware of this in advance.