Meter Gomoku
棋盘是 15×15 的真实对局盘，棋力来自官方 Rapfi 引擎与官方 NNUE 权重。所有分析、评分、胜率、深度、节点都直接取自引擎输出，不做任何二次美化。
The board is a real 15×15 grid, and the strength comes from the official Rapfi engine with official NNUE weights. Every number shown — score, win rate, depth, nodes — is taken straight from the engine, never post-processed.
主按钮直连 GitHub 官方 release 地址下载 Meter.Gomoku_Setup.4.514.exe（44.5 MB）；若所在网络打不开 GitHub，可改用上方蓝奏云镜像。两个来源是同一个安装包。
The primary button downloads Meter.Gomoku_Setup.4.514.exe (44.5 MB) directly from the official GitHub release; if GitHub is unreachable from your network, use the Lanzou mirror above. Both serve the same installer.
01概览Overview
Meter Gomoku 是一个把「强引擎」和「能看清引擎在想什么」放在一起的五子棋工作台。它不只让你和 AI 下棋，更把引擎的搜索过程完整摊开：深度、节点、每秒节点、用时、评估分、胜率、候选点排名，都实时画在棋盘与仪表盘上。
Meter Gomoku pairs a strong engine with full visibility into what that engine is doing. It is not only a place to play — it exposes the whole search: depth, nodes, nodes per second, elapsed time, score, win rate and ranked candidate points, live on the board and dashboard.
另有复盘、识图、连线三类独立窗口
Plus separate Review, Vision and Link windows
禁手判定由引擎完成，与官方口径一致
Forbidden-move logic handled by the engine itself
它适合谁Who it is for
- 想变强的棋手：每一步都能看到引擎的候选点与评分，知道自己差在哪里。
- 做死活题 / 算杀的人：VCF / VCT 前瞻把连续冲四、连续活三的杀法逐手推给你看。
- 开局研究：平衡点用官方成对搜索给出「双方都能接受」的开局点。
- 看别人的棋：识图从截图里读盘，连线则直接盯着屏幕上的对战平台自动应手。
- Players who want to improve — every move shows the engine's candidates and scores, so you can see exactly where you went wrong.
- Puzzle and killing-sequence practice — VCF / VCT look-ahead walks the whole forced win, move by move.
- Opening study — balance points come from the official paired search, giving openings both sides can live with.
- Watching other games — Vision reads a board from a screenshot; Link watches an on-screen client and answers automatically.
02界面导览Interface tour
01 顶栏Top bar
连线 / 识图 / 复盘 / 历史 / 关于 / 设置，六个入口各开独立窗口或抽屉。 Link / Vision / Review / History / About / Settings — six entries, each opening its own window or drawer. |
02 棋盘Board
15×15 交叉点；棋子默认显示手数，分析、防守、平衡与前瞻标记直接画在交点上。 15×15 intersections; stones show move numbers, and analysis, defence, balance and look-ahead marks draw straight on the points. |
03 棋盘下方工具栏Board toolbar
分析计算 · 上一步 / 下一步 · 重新开始 · AI 落一子 · 保存局面 · 旋转 / 翻转 / 平移 · AI 执子。 Analyse · step ← → · reset · AI move · save · rotate / flip / shift · AI side. |
04 对局设置卡Match card
模式、AI 执子、思考时间、后台思考、动态时间、规则，以及线程、哈希、选点范围等引擎信息。 Mode, AI side, think time, background thinking, dynamic time and rules, plus threads, hash, candidate range and other engine info. |
05 引擎仪表盘Dashboard
深度 / 速度 / 节点 / 用时 / 评估 / 最佳，下方是评估曲线、黑白两栏与引擎日志。 Depth / speed / nodes / time / eval / best, followed by the eval curve, the black and white rows and the engine log. |
06 卡片可拖动Cards are draggable
三张卡可以拖到任意一侧、点小横杠折叠、关闭，布局会自动记住。 All three cards can be dragged to either dock, collapsed with the dash or closed; the layout is remembered. |
03对局与使用方法Play & how to use it
默认进入对弈模式（空盘）。点任意交叉点落子；AI 执黑或执白时会自动应手。下面从开始一局、逐项设置到暂停键，把对局流程说明清楚。
The app opens in Play mode on an empty board. Click any intersection to place a stone; the AI answers whenever it owns Black or White. The guide below covers starting a game, every setting and the pause key.
快速开始Quick start
- 在右侧对局设置卡选择模式：对弈 / 摆盘 / 残局。
- 按需要打开AI 执黑或AI 执白；两个都不开＝黑白双方都由你自己下，两个都开＝AI 与 AI 自打。
- 设定思考时间（0.2～300 秒），这是 AI 每手的思考预算；时间越长棋力越强，CPU 占用也越高。
- 在棋盘交叉点点下第一子开始行棋；棋盘下方的AI 执子可以随时改执子方并立即行棋。
- 要清盘就点重新开始（当前这局会先自动存进历史）；要保存当前盘面就点保存局面。
- Pick a mode in the match card on the right: Play / Compose / Endgame.
- Turn on AI plays Black and/or AI plays White. Both off = you play both sides; both on = AI versus AI.
- Set the think time (0.2–300 s): the budget for each AI move. Longer is stronger, at the cost of CPU load.
- Click an intersection to start. The AI side button under the board switches sides on the fly and moves at once.
- Use Reset to clear the board (the current game is saved to History first), or Save to store the position.
对局设置逐项Every match setting
| 设置Setting | 说明What it does |
|---|---|
| 模式Mode | 对弈：AI 自动应答。摆盘：自由摆子，AI 完全不参与。残局：自己摆一个局面来研究；摆好后点「确定」定盘，之后「重新开始」会回到这个局面。Play: the AI answers automatically. Compose: free placement, with no AI involvement. Endgame: build a position to study; confirm it and Reset will return to that position. |
| AI 执子AI side | 两个独立开关：只开AI 执黑＝你执白；只开AI 执白＝你执黑并开第一手；两个都开＝AI 自打；两个都关＝没有 AI。棋盘下方还有一颗「AI 执子」，点开可快速改执子与思考时间，并立即行棋。Two independent toggles. Black only = you play White; White only = you play Black and move first; both on = AI self-play; both off = no AI. The “AI side” button under the board reopens the same choice plus think time, and moves immediately. |
| 思考时间Think time | 0.2～300 秒，内部按毫秒交给引擎。时间越长棋力越强，也会把 CPU 线程吃得更满。0.2–300 s, handed to the engine in milliseconds. Longer is stronger and keeps the CPU threads busier. |
| 后台思考Background thinking | 轮到你思考时，引擎在后台继续深算这一局面；你一落子，它就能更快给出应手。关掉后引擎落子即停，不再空转。While you think, the engine keeps deepening this position, so your move is answered faster and deeper. Turn it off and the engine stops right after its move instead of idling. |
| 动态时间Dynamic time | 开启后不再严格按设定秒数：简单局面提前收手，难局最多多用 15%，见杀局面只用 40%；省下的时间存进时间池，供后续几手调用。关闭后每手严格按设定秒数思考。When on, the engine stops early on simple positions, can spend up to 15% more on hard ones and only 40% on a proven mate; the saved time goes into a pool for later moves. Turn it off to search exactly the set time on every move. |
| 规则Rules | 无禁手：连成五子（含长连）即胜。无禁手 · 长连不赢：正好五子才算胜。连珠（有禁手）：黑方三三、四四、长连均为禁手，判定由引擎完成。换规则会自动重置引擎状态，前瞻与算杀也跟随所选规则。Freestyle: five or more in a row wins. Freestyle, overline loses: exactly five wins. Renju: double-three, double-four and overline are forbidden for Black, judged by the engine. Switching rules resets the engine; look-ahead and VCF/VCT follow the chosen rule. |
| 暂停键Pause key | 棋盘下方中间的 ▶ / ❚❚ 键在三态间循环：AI 走一手＝空闲时点它，引擎立刻替当前行棋方算一手并落下；暂停＝AI 正在思考时点它，立刻停住；自动对弈＝让 AI 连续行棋。The ▶ / ❚❚ key under the board cycles through three states: AI moves once (when idle, search now and play one move), Pause (stop the running search), and Auto-play (let the AI keep moving). |
04计算与分析Analysis
「计算评估」卡片里是全部的分析工具。它们共用一台引擎、一个队列 —— 同一时刻只跑一个任务，落子优先于后台分析，所以你不会看到「算一半被切断」的假分数。
The analysis card holds every tool. They share one engine and one queue — only one job runs at a time, and a real move always pre-empts background analysis, so you never see scores from a half-finished search.
| 按钮Button | 作用What it does |
|---|---|
| 分析计算Analyse | 对当前局面持续深算，把「分析点数」个候选按名次画在棋盘上（圈内是名次与评分）。再点一次＝停止分析，标记保留。Keeps deepening the position and draws the top-N candidates ranked on the board (rank and score inside the circle). Click again to stop; markers stay. |
| 扫描防守Defence scan | 用引擎原生 YXSEARCHDEFEND：假设对手先下，对全盘每个点给出「对手下这里有多危险」，按五档危险度铺成圆环。找急所、找唯一防点非常直观。Native YXSEARCHDEFEND: for every point it reports how dangerous it would be if the opponent played there, drawn as five tiers of rings. Ideal for finding urgent defensive points. |
| 平衡点Balance points | 官方成对平衡搜索（YXBALANCETWO），给出 1～6 个黑方或白方都不吃亏的开局点；空盘固定 3 个（第一手天元）。判「平」的门槛是 |评分| ≤ 50。Official paired balance search (YXBALANCETWO) returning 1–6 opening points neither side can object to; three on an empty board (first move at centre). "Balanced" means |score| ≤ 50. |
| 前瞻 / VCF / VCTLook-ahead / VCF / VCT | 把这之后的一串强制着法摆出来：VCF＝连续冲四取胜，VCT＝冲四加活三的连续威胁。推演子按手数标号，可点代码链预览，确认无误点「确定」落到棋盘。Lays out the forced sequence: VCF wins by continuous fours, VCT by continuous threats mixing fours and open threes. Stones are numbered, code links preview each move, and Confirm plays the line. |
| AI 视图AI view | AI 思考期间在棋盘上铺出它的候选点（圆形，格内是评估分）；AI 一落子就撤掉。While the AI thinks, its candidate points appear on the board (circles with the score inside) and vanish as soon as it moves. |
| 指导视图Guidance view | 轮到你时用方形四色提示最佳落点，你一落子就消失。On your turn, coloured squares suggest the best point; they disappear when you move. |
| 清除计算Clear markers | 只擦掉棋盘上的标记（分析 / 防守 / 平衡 / 前瞻 / VCF / VCT），不动局面本身。Erases every marker (analysis, defence, balance, look-ahead, VCF/VCT) without touching the position. |
| 分析点数Analysis points | 1～8。多点会略慢一点，但能一次看清所有候选；它与对局设置里的「多主变」共用同一份值。1–8. More points is slightly slower but shows every candidate at once; it shares one value with “Multi-PV” in the match settings. |
怎么读仪表盘Reading the dashboard
| 项目Item | 含义Meaning |
|---|---|
| 深度Depth | 当前完成的搜索层数。同一局面、同一硬件下越深越准。Search plies completed so far. Deeper is more accurate on the same position and hardware. |
| 速度Speed | 每秒节点数（nps），反映线程、指令集与当前局面的实际吞吐。Nodes per second (nps), reflecting threads, instruction set and the position’s real throughput. |
| 节点Nodes | 本次搜索累计访问的节点数。Total nodes visited in this search. |
| 用时Time | 本次搜索已用的时间；「分析计算」跨轮会累加显示。Elapsed time for this search; Analyse accumulates it across rounds. |
| 评估Eval | 正＝黑优，负＝白优。+M12 表示黑方 12 手内必胜，-M9 表示白方 9 手内必胜。Positive favours Black, negative favours White. +M12 means Black mates in 12; -M9 means White mates in 9. |
| 最佳Best | 引擎当前的首选落点（棋盘坐标）。The engine’s current first choice, as a board coordinate. |
| 胜率Win rate | 由评估分换算：胜率 ＝ sigmoid(评估 ÷ 200)。黑方 88% 与白方 12% 是同一局面的两面。Derived from the score: win rate = sigmoid(eval ÷ 200). Black 88% and White 12% are the two sides of one position. |
| 评估曲线Eval curve | 整局评分走势：黑线黑方、蓝线白方，纵轴随分值自动进位（最大 ±1200）；浮到曲线上可读每一手的分数。Score history for the game: black line for Black, blue for White, with the axis scaling automatically (up to ±1200). Hover a point to read that move’s score. |
| 黑 / 白两栏Black / White rows | 分别给出该方的分数与胜率，方便对照同一局的两种视角。Score and win rate for each side, so both views of the same position can be read together. |
05棋盘与局面Board & position
| 操作Control | 说明Detail |
|---|---|
| ← / → | 回看上一手 / 下一手。回看时引擎状态不变，回来还是原来的局面。Step back / forward. Stepping does not disturb the engine. |
| 重新开始Reset | 清空棋盘并保留所有设置；清盘前会先把当前这局自动存进历史。残局模式下定过盘，则回到那个已确定的局面。Clears the board and keeps every setting; the current game is auto-saved to History first. In Endgame mode it returns to the confirmed position. |
| 保存局面Save | 存进「历史」，可多选、可导入导出、可标记为永久。Stores into History — multi-select, import/export, mark as permanent. |
| AI 落一子 / 暂停AI move / pause | 棋盘下方中间的 ▶ / ❚❚ 键：空闲时让引擎替当前行棋方算一手并落下，AI 思考中则暂停；三态循环见第 03 节。The ▶ / ❚❚ key under the board: when idle it makes the engine search and play one move for the side to move; while thinking it pauses. See section 03 for the three-state cycle. |
| ↻ ⇄ ✥ | 整体旋转 90°／左右·上下·对角翻转／整局平移一格。坐标轴不动，方便对照棋谱。Rotate 90°, mirror horizontally / vertically / diagonally, or shift the whole position one cell. The axes stay put so coordinates remain comparable. |
| 局面代码Position code | 与 gomocalc 同格式：每手两个字符，列字母 + 行号，例如 h8h9g7f7i8。可粘贴载入，也可复制分享。gomocalc-compatible: two characters per move, column letter + row number, e.g. h8h9g7f7i8. Paste to load, copy to share. |
摆盘 / 残局模式Compose & endgame
摆盘模式自由摆子，AI 完全不参与。残局模式用于「自己摆一个局面来研究」，棋盘下方会出现「确定 / 顺序摆盘 / 任意摆盘」：顺序摆盘＝黑白轮流落子；任意摆盘＝先选「黑子」或「白子」，再连续摆这一色，数量与顺序都不限。摆好后点「确定」定盘，按键变绿并显示「已确定」；再点一次解除，可以重新摆。残局模式下 AI 全程不参与，交换手类规则自动变灰。
Compose mode is free placement with no AI at all. Endgame mode is for building a position to study: the board toolbar gains Confirm / Sequential / Free. Sequential alternates Black and White; Free lets you pick Black or White and place that colour any number of times in any order. Confirm locks the position (the key turns green and reads “Confirmed”); press it again to release and keep editing. The AI stays out of Endgame mode, and swap-related rules are greyed out.
06复盘Review
顶栏「复盘」打开独立的复盘窗口，专为「把一盘棋读透」设计：
The Review window from the top bar is built for reading a game closely:
- 背诵复盘：隐藏后面的手，凭记忆一手一手摆出来。摆错会立刻标红并计数，可以随时「重来」。
- 回顾复盘：以 1 秒/手自动播放，速度可调；可开「序号」看手数，也可以随时跳到任意一手。
- 复盘窗与主窗互不干扰：复盘只读局面，不会改动你正在下的棋。
- Recite: later moves are hidden and you replay the game from memory. A wrong stone is flagged and counted; restart any time.
- Replay: auto-plays at one move per second (adjustable), with optional move numbers and jump-to-move.
- The review window reads the position only — it never disturbs the game in the main window.
07识图Vision
顶栏「识图」把一张图片或一段屏幕画面变成可以继续下的局面。
Vision turns a picture — or a piece of your screen — into a position you can keep playing.
- 上传图片 / 屏幕截图，然后裁剪到只剩棋盘。裁得越紧，识别越稳。
- 点识别，等待结果。识别参数（5 个阈值）在面板里可调，拖动即时生效并自动保存。
- 核对棋盘上的棋子，确认后加载到练习：局面直接进入摆盘模式，可以继续分析或对弈。
- VC 模式允许黑白子数不统一（残局、摆拍题常见），加载后自动跑一次 VCF/VCT。
- Upload or capture, then crop to the board only. Tighter crops recognise more reliably.
- Press Recognise. The five thresholds in the panel are live and auto-saved.
- Check the stones, then load into practice — the position lands in Compose mode, ready to analyse or continue.
- VC mode accepts uneven stone counts (common in puzzles) and runs a VCF/VCT pass after loading.
08连线Link
连线是「替你在别的程序里下棋」的功能：它不接管对方的程序，而是盯着屏幕。
Link plays for you inside another program — not by hooking it, but by watching the screen.
- 顶栏「连线」打开连线窗，点框选棋盘，在屏幕上拖出目标棋盘的范围。
- 点识别。框从紫色变成蓝绿色、四角出现 L 形角框＝识别成功。
- 选AI 执黑或AI 执白（互斥），设好思考时间。这些设置会同步回主窗。
- 识别在后台静默进行：画面静止时几乎零开销，棋子一变才重新读盘，读到你的回合就自动落子。
- Open Link, press Select board, drag a rectangle over the target board.
- Press Recognise. The frame turns from purple to teal with L-shaped corners = locked on.
- Pick AI plays Black or AI plays White (mutually exclusive) and a think time; both sync back to the main window.
- Recognition runs silently: nothing when the image is static, a re-read only when a stone appears, and a move as soon as it is the AI's turn.
09引擎与硬件Engine & hardware
棋力全部来自 Rapfi。软件把「能调的」和「不该动的」分开了：线程数、哈希、变体、权重、选点范围可以调；搜索的核心参数一律采用官方配置。
All strength comes from Rapfi. The app separates what you may tune from what should be left alone: threads, hash, variant, weights and candidate range are yours; the core search parameters stay at the official values.
| 项Item | 说明Detail |
|---|---|
| 线程数Threads | 默认＝物理核数（本机 8 核 → 8 线程）。物理核跑满比超线程更稳，评测抖动更小。Defaults to the physical core count (8 on an 8-core machine). Filling physical cores is steadier than over-subscribing with SMT. |
| 哈希表Hash | 置换表大小，越大越省重复搜索。默认按内存自适应，32 GB 机器给 1～2 GB 都合理。Transposition table size; bigger avoids re-searching. It adapts to your RAM by default; 1–2 GB is reasonable on 32 GB. |
| 变体Variant | sse / avx2 / avx512 / avx512vnni / avxvnni 五个二进制，启动时按 CPU 指令集自动挑，启动日志里会写「本机实测最快」。Five binaries (sse / avx2 / avx512 / avx512vnni / avxvnni), picked at startup from your CPU features; the log states which one was chosen. |
| 权重Weights | NNUE（默认，判断力强）或传统估值（更省算力）。NNUE (default, strongest judgement) or the classical evaluator (cheaper). |
| 选点范围Candidate range | 0 最窄（5×5）～ 5 全盘。窄＝同时间算得更深；宽＝不容易漏点。默认 3（7×7+4 线）是官方推荐档。0 (5×5) … 5 (whole board). Narrower searches deeper in the same time; wider misses fewer points. Default 3 (7×7 + 4 lines) is the recommended step. |
| 最大深度 / 节点Max depth / nodes | 上限保护：最大深度 4～128，最大节点 0.1～4096 M。留空＝由思考时间决定。Safety caps: max depth 4–128, max nodes 0.1–4096 M. Leave them empty to let think time decide. |
10设置与主题Settings & themes
顶栏「设置」打开独立的设置窗，它与主窗双向实时联动：改线程、哈希、选点范围、分析点数、主题、语言，所有窗口立刻跟着变。
The Settings window is a separate window that stays in sync both ways: threads, hash, candidate range, analysis points, theme and language apply to every window immediately.
主题Themes
20 套预设（浅色、深色 + 18 套色系），或「自定义」：棋盘与背景分别调色，HSV 色板、RGB 数值、预制色、一键恢复默认。
20 presets (light, dark and 18 colourways) or a custom mode: board and page background coloured separately with an HSV pad, RGB fields, presets and a reset.
字体Fonts
界面字体、引擎日志字体、棋子数字字体分别可选，字号也可缩放。字体列表带实时预览。
UI font, engine-log font and the stone-number font are chosen separately, with scalable sizes and live previews.
语言Language
中文 / English 一键切换，所有窗口同步，下次启动保持。
Chinese / English with one click, applied to every window and remembered next launch.
11常见问题FAQ
分数为什么和别的软件不一样？Why does the score differ from another program?
评估分是「相对当前行棋方」的原始分，不同引擎、不同阈值、不同搜索深度都会不同。判断强弱看胜率和最佳点更可靠；分数每次刷新略有跳动也正常，那是搜索还没收敛。
The score is a raw value relative to the side to move, so different engines, thresholds and depths all differ. Win rate and the best point are more reliable; small jitter between refreshes is normal before the search converges.
引擎没反应 / 一直「引擎还没就绪」Engine unresponsive / "engine not ready"
确认 rapfi-native 与权重文件还在 exe 旁边；路径不要有中文；被杀毒拦过就先放行再重启；内存吃紧时把哈希调小一档。
Check that rapfi-native and the weights still sit next to the exe; avoid non-ASCII paths; whitelist it if antivirus interfered; drop the hash one step if RAM is tight.
风扇很响、CPU 一直 100%Loud fans, CPU pinned at 100%
只要还有搜索在跑，引擎就会占满线程。关掉「分析计算」就恢复空闲；对弈只在 AI 回合搜索。
As long as a search is running the engine fills its threads. Turn off Analyse and it idles; during a game it only searches on the AI's turn.
换规则 / 开新局后评分很怪Odd score after switching rules or starting over
换规则会自动清空置换表并重置引擎；如果仍有异常，点一次「重新开始」，让第一手在干净状态下重新算。
Switching rules clears the hash and resets the engine. If something still looks off, hit Reset so the first move is searched from a clean state.
为什么有的候选点胜率都是 100%？Why are some candidates all at 100%?
说明引擎已经确认了必胜或必败（评分显示 +M12 这种杀棋值）。这时排名意义不大，看杀棋手数就行。
The engine has proven a forced win or loss (the score shows a mate value such as +M12). Ranking matters less then — read the mate distance instead.
12Meter Web · 自研网页版Meter Web · our own web build
如果你现在正用手机，不妨体验我们自研界面的网页版五子棋计算器（Meter Gomoku web）——它把桌面版那套仪表盘搬进了浏览器：深度、速度、节点、用时、胜率条、分析路线一屏可见，界面语言、棋子样式、思考时间、线程数都能自己调，算完直接落子。
If you are on a phone right now, try Meter Gomoku web — our own web calculator. It brings the desktop dashboard into the browser: depth, speed, nodes, time, the win-rate bar and the analysed line on one screen, with your own choice of language, stone style, think time and thread count — and it can play the move it just calculated.
棋盘在上、仪表盘在下并铺满整宽；设置收进底部抽屉，横屏时自动切成左仪表盘、右棋盘两栏
Board on top, full-width dashboard below; settings live in a bottom sheet and landscape switches back to dashboard-left, board-right
Rapfi NNUE 编译成 WebAssembly，权重建在本机缓存里，断网也能复盘
Rapfi NNUE compiled to WebAssembly; the weights are cached locally so it keeps working offline
没有对弈、复盘、识图、连线——那些仍然交给桌面版；这里只有摆盘、计算、平衡点与落子
No play, review, vision or Link — those stay in the desktop app. Here it is compose, analyse, balance points and play
—关于About
Meter Gomoku 由个人开发与维护，界面与引擎调用层均为自研；棋力引擎与权重来自开源项目 Rapfi（作者 dhbloo，GitHub 账号现显示为 dotb1ue；项目主页见下方「Rapfi · 引擎源码」）。如果你发现界面或引擎调用上的问题，欢迎在项目仓库提 issue。
Meter Gomoku is an independently built and maintained desktop app; the UI and the engine-calling layer are original, while the engine and weights come from the open-source Rapfi project (author dhbloo, whose GitHub account now appears as dotb1ue; the project page is linked below as “Rapfi · engine source”). Bug reports are welcome in the project repository.