#译脉·先知 2.0 (MoonBit)

一套带预测能力的记忆网络——让本地智能体记住工作流，并在遇到同类任务时预测下一步需求、给出可白盒解释的路径。这是「译脉·先知 2.0 预知记忆网络」引擎的 MoonBit 零依赖实现（仅 core/json + core/math）。

在「预测记忆」内核之外，已落地 #22 翻译记忆（TM）/ 术语库（TB）一等公民：真正的 fuzzy match（含匹配率%）、concordance 检索、TBX 术语库强制对齐与一致性校验——让引擎从「只预测」走向「预测 + 检索 + 术语守门」。

📜 变更历史：CHANGELOG.md 记录每次 P 增量的完整 commit 列表与影响范围；本 README 只描述当前状态。

#Table of Contents

• Features
• How it works (D1–D8)
• Installation
• Quick Start
• Modern Corpus Evaluation (2025–2026)
• Service Layer — 纯 MoonBit HTTP API (cmd/service)
• API Reference
• #22 TM/TB — Translation Memory & TermBase
• Data Formats
• Evaluation & Test Results
• Roadmap & Extension Status
• For Agents / Integration
• Contributing
• License
• Acknowledgements

#Features

• Predictive memory network — remembers workflows and predicts the most likely next step before the user asks.
• White-box explainability — every prediction/explain returns the concrete node/edge/transition path, never a black box.
• Semantic recall — given a new source sentence, activates a spreading network to recall the exact bilingual terms/sentences that keep terminology consistent.
• Cold-start generalization — role-level abstraction (D8) induces domain-independent process skeletons, so an unseen project still gets a sensible next step.
• Deterministic & reproducible — a logical clock replaces wall-clock time; identical call sequences yield byte-identical to_json output (no RNG).
• Zero third-party dependencies — pure MoonBit core (json + math only); nothing to install beyond the moon toolchain.
• Serializable — full engine state exports/imports as JSON for persistence and cross-session restore.
• Fast inference (快) — role inverted-index (role_members) + per-source Top-8 pruning keep the hot path off full-graph scans; a pred_cache (fully invalidated on any engine change, not LRU) short-circuits repeated (context, k) queries.
• Accurate (准) — second-order Markov (trans2, P(w3|w1,w2) blended at λ=0.4), multi-granularity role keys (前二/前四/前后各二), elastic forgetting (recency-aware edge decay), adaptive Hebbian LR, per-domain bias ΔW (LoRA-style), online contrastive learning (cl_step), and attention-gated edge weights in recall.
• Explainable & bilingual (美) — TermNode (mark_term) boosts terminology recall with a +5.0 activation and a "term hit" flag; explain_card returns a white-box activation_path / prediction_path / value_breakdown JSON; align_diff gives a character-level LCS edit script for bilingual alignment.
• TM / TermBase (检索 + 守门) — #22 新增 add_tm / fuzzy_match / concordance / load_tbx / enforce_terms / check_terms：真正的 fuzzy match（匹配率 %）、concordance 检索、TBX(ISO 30042) 术语库解析、术语强制对齐与一致性校验（见 §#22）。
• Incremental & collaborative (中/长周期) — Write-Ahead Log (wal_*) for event-sourced replay, active-learning candidates by uncertainty + diversity, and federated increment export/import (fed_*) for cross-agent coordination.
• Pure-MoonBit service layer (纯 MoonBit 全栈) — cmd/service 起本地 HTTP server（127.0.0.1:8787），27 个 /api/* 端点 + /mcp MCP Server（13 基础 + retrieve_prompt / bleu / chrf / style_check / style_report / back_align / term_conflicts / fed_export / fed_import / distill_inject / active_learning / metrics / health / mqm_re_annotate；MCP 层共 25 tools）全部实测通过，含 BLEU/chrF++ 评测、风格检查、回译对齐、术语冲突、TMPlm、联邦/蒸馏注入、主动学习推荐；TM 状态经 fs 原子写持久化（tmp+rename），重启恢复闭环——引擎到服务零桥接语言。

#How it works (D1–D8)

#Project Structure

yimai_prophecy_moonbit/
├── engine.mbt                  # Layer0: 零依赖预测记忆引擎内核（D1–D8, #22 TM/TB, MQM）
├── util.mbt                    # 编码/TF-IDF/对齐/URL解码工具函数（P4 decode_pct 新增）
├── yimai_prophecy_moonbit.mbt  # Lib 主入口（routes_meta 单一源 + lib 共享 helper 文档化）
├── tests/                      # P5 仓库整理：18 个测试按主题分到 3 个 sub-package
│   ├── core/                   #   核心/经典测试（69 测试）
│   │   ├── moon.pkg            #   sub-package（独立 wasm-gc 测试目标）
│   │   ├── _test_helpers.mbt   #   跨子包共享 helper（canon/topics + 6 fn，pub）
│   │   ├── yimai_prophecy_moonbit_test.mbt          # 主测试（L1–L2 批量 Hit@3）
│   │   ├── yimai_prophecy_moonbit_accept_test.mbt   # 验收测试台
│   │   ├── yimai_prophecy_moonbit_bench_p1.mbt      # 基准 P1（pruned vs full acc）
│   │   ├── yimai_prophecy_moonbit_benchmark_test.mbt
│   │   ├── yimai_prophecy_moonbit_golden_test.mbt    # 黄金集回归
│   │   ├── yimai_prophecy_moonbit_long_text_test.mbt # 长文本 / 分段 / 数字 token
│   │   ├── yimai_prophecy_moonbit_tm_test.mbt       # TM 专项
│   │   ├── yimai_prophecy_moonbit_v2_test.mbt       # V2 引擎 API
│   │   └── yimai_prophecy_moonbit_wbtest.mbt        # 白盒内部测试
│   ├── corpus/                 #   语料/数据集驱动测试（54 测试）
│   │   ├── moon.pkg
│   │   ├── _test_helpers.mbt   #   inline 副本（与 core/ 同步）
│   │   ├── yimai_prophecy_moonbit_extended_corpus_test.mbt    # 6 个领域 11 测试
│   │   ├── yimai_prophecy_moonbit_modern_corpus_test.mbt      # Modern Corpus：8 个前沿领域
│   │   ├── yimai_prophecy_moonbit_roadmap_test.mbt            # Roadmap 增量语料
│   │   ├── yimai_prophecy_moonbit_frontier_corpus_test.mbt    # Frontier Corpus：10 个领域 14 测试
│   │   └── yimai_prophecy_moonbit_business_corpus_test.mbt    # 商务领域（P5 新增，ISO 11669 / GB/T 30539）
│   └── feature/                #   扩展/新功能测试（52 测试）
│       ├── moon.pkg
│       ├── yimai_prophecy_moonbit_extension_test.mbt          # 扩展能力回归（E1–E18）
│       ├── yimai_prophecy_moonbit_quality_test.mbt            # P4 质量/安全（T30–T37）
│       ├── yimai_prophecy_moonbit_routes_test.mbt             # 端点元数据单一源测试
│       └── yimai_prophecy_moonbit_mqm_reannotation_test.mbt   # P5 MQM 二次标注（Google 2025-10-28）
├── cmd/
│   ├── main/moon.pkg          # demo 程序（训练 → Hit@3 → replay → D8 冷启动 → consolidate → reward → restore）
│   └── service/
│       ├── moon.pkg           # 服务入口（`moon run cmd/service --target native` → 127.0.0.1:8787）
│       ├── mcp.mbt            # MCP Server 实现（spec 2025-11-25 Streamable HTTP）
│       ├── routes.mbt         # 27 个 HTTP 端点路由（24 + metrics + health + mqm_re_annotate）
│       ├── tm_store.mbt       # 引擎持久化（`save_store` 深度守卫 P4）
│       └── web/               # 前端工作台（静态资源，`serve_static` URL 解码 P4 修复）
├── scripts/
│   ├── dev.ps1                # 一键起服务（env check → build → run → seed → smoke）
│   ├── push.ps1               # 双 remote 推送（github via ghproxy.net + gitlink）
│   ├── smoke.ps1              # 烟雾测试（27 端点 + MCP）
│   └── reorganize_repo.py     # 仓库整理复现脚本（git mv + add @lib. prefix + sub-package init）
├── .githooks/
│   └── pre-commit             # 本地门禁（`moon check` + `moon test --target wasm-gc`）
├── docs/
│   ├── skill/SKILL.md         # WorkBuddy 技能编排手册（frontmatter agent_created=true）
│   ├── harness-configs/       # 13 个 harness 配置（Claude Code / Cursor / Gemini CLI / ...）
│   ├── plans/                 # 项目级 plan / note（按日期 YYYY-MM-DD-<topic>.md 命名）
│   └── roadmap.md             # 项目路线图（中文转英文，仓库国际友好）
├── AGENTS.md                  # AI agent 集成指南（含 Project layout 段：未来 _test.mbt 必须在子目录）
├── README.md                  # 项目说明（badge 175/175 + P6 hardening + International Standards）
├── CHANGELOG.md               # 版本变更记录
└── LICENSE                    # MIT License

• Layer0（内核）：engine.mbt + util.mbt —— 零依赖，仅依赖 moonbitlang/core/json + core/math
• Layer1（服务层）：cmd/service/* —— 纯 MoonBit HTTP/MCP 服务，27 个端点，原子写持久化
• Layer2（知识层）：docs/* + AGENTS.md + SKILL.md —— 文档 + 技能编排 + 集成指南

• 契约回归：175 个 test 跨 19 个文件（3 sub-package：core 69 / corpus 54 / feature 52）
• 门禁：scripts/dev.ps1 + .githooks/pre-commit + .github/workflows/ci.yml —— 构建/提交前自动运行 moon test --target wasm-gc
• 目标：175/175 全绿（wasm-gc 目标，可复现）

#Installation

moon add Across2005/yimai_prophecy_moonbit

import {
  "Across2005/yimai_prophecy_moonbit" @lib,
  "moonbitlang/core/json" @json,
}

Requires the MoonBit toolchain (moon, v0.1.2026+).

#Out-of-the-box setup (Windows, AI agents welcome)

powershell -ExecutionPolicy Bypass -File scripts/dev.ps1

确定性回归门禁（纯本地，零云端依赖）: scripts/dev.ps1 在构建前自动运行 moon test --target wasm-gc（175/175 契约回归，P6 hardening 后），任何一项失败即中止。 此外，仓库自带本地 pre-commit hook（.githooks/pre-commit，moon check + moon test --target wasm-gc），已通过 git config core.hooksPath .githooks 接入本仓库——每个 commit 前自动挡住破坏确定性契约的改动。

Windows native prerequisites: cmd/service requires MSVC (link.native.cc in cmd/service/moon.pkg and cmd/main/moon.pkg points at cl.exe — update both if the path differs on your machine); after a MoonBit toolchain upgrade, rebuild the core native bundle once (cd ~/.moon/lib/core && moon clean --target-dir _build/native &&moon bundle --target native --release). AI agents: see AGENTS.md for the full out-of-the-box guide.

#Quick Start

pub fn quickstart() -> Unit {
  let mut eng = @lib.ProphecyEngine::make()

  // 1) observe() records real steps in order; the engine maintains a
  //    context window and a 1st/2nd-order Markov transition model internally.
  let _ = eng.observe("解析源文件结构", "step")
  let _ = eng.observe("提取核心术语表并锁定", "step")
  let _ = eng.observe("生成双语对照草稿", "step")

  // 2) predict Top-3 most likely next steps from current context.
  let pred = eng.predict(3)
  println(@json.stringify(pred))

  // 3) recall: given a query, return related memories (with activation + path).
  let hits = eng.recall("术语", 5)
  println(@json.stringify(hits))

  // 4) persist & restore.
  let snap = eng.to_json()
  let eng2 = @lib.ProphecyEngine::from_json(snap)
  let _ = eng2

  // 5) #22 — TM / TermBase: add memory, load a TBX glossary, align & verify.
  let _ = eng.add_tm("电池包热管理策略", "Battery pack thermal management strategy")
  let tbx =
    "<martif><text><body>" +
    "<termEntry id=\"1\"><langSet xml:lang=\"en-US\"><ntig><termGrp><term>sensor</term></termGrp></ntig></langSet>" +
    "<langSet xml:lang=\"zh-CN\"><ntig><termGrp><term>传感器</term></termGrp></ntig></langSet></termEntry>" +
    "</body></text></martif>"
  let _ = eng.load_tbx(tbx)
  let tmx = eng.fuzzy_match("电池包热管理", 3, 0.70)   // Top-K with match_pct
  let v  = eng.check_terms("install the sensor", "安装设备")  // 1 violation (漏译 传感器)
  println(@json.stringify(tmx))
  println(@json.stringify(v))
}

observe(text, mtype) builds edges/transitions from the current context window and advances the logical clock. To control co-occurrence manually, call remember(text, mtype, ctx) with an explicit ctx array.

moon build
moon run cmd/main     # training → Hit@3 → replay prediction → D8 cold-start → consolidate → reward → restore

#Start the HTTP service (pure MoonBit, cmd/service)

moon build cmd/service --target native && moon run cmd/service --target native
# 译脉引擎服务: http://127.0.0.1:8787

curl -X POST localhost:8787/api/add_tm        -d '{"src":"电池包热管理策略","tgt":"Battery pack thermal management strategy"}'
curl -X POST localhost:8787/api/fuzzy_match   -d '{"query":"电池包热管理方案","k":3,"threshold":0.5}'
curl -X POST localhost:8787/api/check_terms   -d '{"source":"install the sensor","target":"安装设备"}'
curl -X POST localhost:8787/api/qe_auto       -d '{"source":"a","target":"b","match_rate":0.8}'
curl -X POST localhost:8787/api/predict       -d '{"k":3}'

#Modern Corpus Evaluation (2025–2026)

#Corpus domains (Modern + Extended + Frontier)

#Evaluation results — 25 tests, all passing (Modern + Extended + Frontier)

Engine-wide Hit@3 = 0.7773. All 8/8 domains produce valid, domain-specific workflow predictions.

• fuzzy_match("RAG chunking vector store optimization") → hits AI/NLP entries with sim_token / sim_tfidf / sim_char / sim_ngram / sim_tokenset breakdowns
• fuzzy_match("CRISPR knockout of PCSK9 gene") → hits CRISPR domain entries
• fuzzy_match("neural decoding of brain signals") → hits neuroscience content via semantic overlap (threshold 0.15)

• "LLM safety benchmarking" → activates AI Safety + NLP domains
• "gene editing + neural decoding" → activates CRISPR + Neuroscience domains
• "quantum + distributed consensus" → activates Quantum Computing + Distributed Systems domains

#Summary

#Extended Corpus Evaluation — 6 New Frontier Domains

#Additional domains

#Evaluation results — 14 tests, all passing

#Frontier Corpus Evaluation — 10 Emerging Domains

#Frontier domains

#Evaluation results — 14 tests, all passing

• 超长上下文预测：每个 Frontier domain 训练 10+ 步超长工作流，验证 D3 Forward model 在深度上下文下的稳定性
• 复杂事实推理：哲学/数学领域的抽象推理路径测试，验证 D1–D8 记忆网络在事实密集型场景的保持
• 跨学科召回：测试 "AI Safety × Philosophy" 混合查询是否能正确激活两个领域的记忆节点
• 指令/事实混合模式：Frontier Corpus 采用 instruction-fact 混合模式（add_tm 存储事实，observe 学习指令），更贴近真实工作流

#Service Layer — 纯 MoonBit HTTP API (cmd/service)

#架构演进：旧架构 vs 新架构

演进动机：旧架构的引擎能力只能被「会 MoonBit 的程序」消费；新架构让任何会 HTTP 的宿主（浏览器前端、Agent 工具调用、LLM 函数调用）都能用上确定性记忆引擎——内核零依赖铁律不变，只是多了一层纯 MoonBit 的 I/O 壳。

#端点矩阵（27 个，全部 curl 实测通过）

POST /api/fuzzy_match  {"query":"电池包热管理方案","k":3,"threshold":0.5}
→ [{"id":"m3","source":"电池包温度管理方案","target":"Battery pack temperature management plan",
    "score":0.7377,"match_pct":73.7723,"sim_token":0.7125,"sim_tfidf":0.6796,
    "sim_char":0.7778,"sim_ngram":0.6667,"sim_tokenset":0.75}, ...]

#三层关联与记忆闭环

前端操作 ──HTTP──▶ Layer2 服务端点 ──调用──▶ Layer0 引擎方法
   ▲                                              │
   │                                              ▼
   └─── JSON 响应（白盒分数/证据链）◀── save_store() 原子落盘 ◀┘
                                │
                                ▼
                  重启 load_store() → from_json() → 记忆不丢

• 采纳闭环：前端「采纳译文」→ /api/reward{mid,+1} → predictive_value 提升 → 下次 /api/predict 排序更优；
• 学习闭环：前端「记录步骤」→ /api/observe{text} → 转移计数 → 落盘 → 重启恢复 → 预测更准（实测：observe 两步 → predict Top1 prob=1.0）。

#MCP Server（/mcp 端点）—— 供 Claude Desktop / 通用 MCP 客户端消费

# MCP 握手（curl 模拟客户端）
curl -X POST localhost:8787/mcp -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}'
curl -X POST localhost:8787/mcp -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'
curl -X POST localhost:8787/mcp -d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"fuzzy_match","arguments":{"query":"电池","k":2}}}'

协议：initialize（protocolVersion 2025-11-25 + capabilities.tools）/ notifications/initialized（202）/ tools/list（25 tools，inputSchema JSON Schema 2020-12）/ tools/call（未知工具 -32602，引擎异常 isError:true）；GET /mcp 回 405。实现为自建轻量 JSON-RPC 2.0 层（cmd/service/mcp.mbt，MoonBit 无现成 MCP 库），全程复用引擎 @lib.obj/str_json/num_json 构造（Json 为 FFI 类型）。

#持久化

• 引擎状态经 to_json() → @fs.write_file(tmp, create_mode=CreateOrTruncate) → rename 原子落盘到 tm_store.json；
• 重启时 load_store() → from_json() 完整恢复（实测 tm_count 持久化后重启一致）；
• 单例持有：engine_ref : @ref.Ref[@lib.ProphecyEngine]（MoonBit 顶层无全局可变变量，Ref 是标准方案）。

#平台要求（Windows）

• async 的 native 后端硬性要求 MSVC（thread_pool.c: #error "Currently only MSVC is supported on Windows"），mingw gcc 不可用；wasm/js 后端暂不支持 socket server；
• 构建需 MSVC 环境（INCLUDE/LIB）+ moon.pkg 配置 link.native.cc 指向 cl.exe；
• 工具链升级后需重建 core native bundle（cd ~/.moon/lib/core && moon clean --target-dir _build/native && moon bundle --target native --release）。

#API Reference

#Core (D1–D8, persistence, feedback)

#Learning & explanation (准·美)

#Incremental & collaborative (中/长周期)

##22 TM/TB (检索 + 术语守门)

##22 TM/TB — Translation Memory & TermBase

• add_tm(src, tgt) builds a MemoryNode of mtype="tm" carrying text=source, translation=target (and maintains the TM document-frequency index for IDF).
• fuzzy_match (S1 upgrade) scores 0.55·idf_dice + 0.20·char-2gram-dice + 0.15·token-set-dice + 0.10·position over mtype=="tm" nodes only. The IDF table (ln((N+1)/(df+1))+1) down-weights frequent words so rare domain terms dominate (R23); char 2-gram catches morphological variants; token-set Dice tolerates word reordering — a Chinese reordered query scores 0.91 with the new formula while the legacy one misses it entirely (R24). The old 0.7·token-cosine + 0.3·char-ratio formula remains as fuzzy_match_legacy for A/B comparison.
• concordance counts query-term occurrences per TM segment — concordance % is term occurrence, distinct from the fuzzy similarity score (per the research baseline).
• load_tbx parses TBX 2.0 (<ntig><termGrp><term> or simplified <tig><term>), language-aware via xml:lang, into mtype="term" nodes with is_term=true.
• enforce_terms / check_terms provide terminology lock-in and missed-term detection, with word-boundary-aware matching for Latin terms.

#Data Formats

#Engine persistence (to_json / from_json)

{
  "memories": {
    "m1": { "id":"m1","text":"解析源文件结构","type":"step",
            "vec":{"解析":1,"源":1},"created":1,"last_used":3,
            "use_count":2,"feedback":0,"edges":{"m2":0.30},
            "predictive_value":0.42,"hit_count":1,"predict_count":1,
            "is_term":false,"translation":"" }
  },
  "transitions": { "m1": {"m2":1.0} },
  "episodes": [["m1","m2","m3"]],
  "context": ["m1","m2","m3"],
  "stats": { "preds":1, "hits":1, "remembers":3, "evolutions":0 },
  "seq": 3, "explore": 0.0, "clock": 3, "meta_hits": [1],
  "snapshot": {}, "role_trans": {}, "role_index": {}
}

TM / TermBase nodes add two fields: a TM node carries "type":"tm","translation":"<target>"; a terminology node carries "type":"term","is_term":true,"translation":"<target term>". Both are round-trip preserved through to_json/from_json (covered by R20).

#predict(k) returns

{
  "predictions": [
    { "id":"m4", "text":"生成双语对照草稿", "prob":0.62,
      "path": [ { "from":"m3", "p":0.55 } ] }
  ],
  "confidence": 0.62,
  "uncertainty": 0.41
}

#recall(query, k) returns (array)

[
  { "id":"m2", "text":"提取核心术语表并锁定", "type":"step",
    "score":0.71, "activation":1.0, "via_edges": [ { "to":"m1", "w":0.30 } ] }
]

#consolidate(prune) returns

{ "pruned": 0, "nodes": 3, "edges": 2, "explore": 0.0, "recent_hit_rate": 0.5 }

##22 — TM / TermBase outputs

[
  { "id":"m7", "source":"电池包热管理策略", "target":"Battery pack thermal management strategy",
    "score":0.91, "match_pct":91.0, "sim_token":0.90, "sim_tfidf":0.88,
    "sim_char":0.95, "sim_ngram":0.86, "sim_tokenset":0.93 }
]

[
  { "id":"m11", "source":"打开设置菜单选择网络", "target":"Open Settings menu, choose Network", "hits":1 }
]

<martif><text><body>
  <termEntry id="1">
    <langSet xml:lang="en-US"><ntig><termGrp><term>network logon</term></termGrp></ntig></langSet>
    <langSet xml:lang="zh-CN"><ntig><termGrp><term>网络登录</term></termGrp></ntig></langSet>
  </termEntry>
</body></text></martif>

[
  { "term":"network logon", "translation":"网络登录", "mid":"m20" }
]

[
  { "term":"sensor", "expected":"传感器", "mid":"m21" }
]

#Evaluation & Test Results

#Credibility hardening

• Classic acceptance suite (4 tests): Layer1 batch training → Hit@3=0.8246, determinism, JSON round-trip, consolidation.
• Modern corpus suite (11 tests): 8 cutting-edge English domains, 240 training observations, Hit@3=0.7773, cold-start generalization on unseen domains, cross-domain semantic recall, deep fuzzy match, attention-gated recall, WAL event sourcing, federated export/import, and distillation.
• Regression suite (R1–R25): TM/TB fuzzy match % (S1 IDF + 2-gram + word-order), concordance, TBX load+enforce+check, word-boundary (log≠logical), 3-language xml:lang, IDF discrimination, word-order tolerance, empty/short-query boundary.
• Extension suite (E1–E18): QE+MQM, format-fidelity, multimodal-OCR-stub, batch-CI, TMS XLIFF/TMX, observability/drift.

cd yimai_prophecy_moonbit
moon test --target wasm-gc      # all 175 tests (P6 hardened)
moon test --target wasm-gc --filter Layer*   # modern + extended + frontier corpus (25 tests)
moon test --target wasm-gc --filter T*       # P4 quality/security tests (T30–T51, 22 tests)
moon build --target wasm-gc     # library only
cd cmd/main && moon build --target wasm-gc && moon run .

#Extension API (#2–#7)

Boundary principle. Capabilities #4 (OCR) and any host persistence stay outside the zero-dependency engine. The engine speaks JSON at these boundaries, so the host (Node/Python/Agent) supplies OCR, files, and I/O — the MoonBit core stays 100% pure-stdlib and wasm-gc-testable.

#Roadmap & Extension Status

#Engine implementation status (内核「快·准·美」)

#Seven extension capabilities (user-scoped)

This project is not packaged as a WorkBuddy skill yet. The dev loop for these is: research (Deep Research / WebSearch) → review (open-code-review) → verify (browser automation + moon test). MoA is intentionally not embedded inside the skill (kept as an external advisor).

#International Standards & Compliance (P4 增量)

2026-08 增补：随着项目演进，本节列出当前已声明对齐 / 仍属 roadmap 的国际/区域标准， 以及对应的本地化合规姿态。yimai 本身是技术构建块（library + local service）而非翻译服务 机构；本节为「adopter 集成指南」，非 ISO 认证声明。

#已对齐（algorithm / docs 层）

#MQM 严重度尺度（与业界三方对齐，P5 增量）

⚠️ mqm_re_annotate 当前实现是 deterministic 自重审：因 qe_auto 算法本身确定性， re_severity == original_severity 恒成立，consistent: true。 端点接口与 JSON schema（re_annotated / critical_count / re_annotations 三段式）已 按 Google 2025-10-28 论文语义设计；多标注员模型（multi-rater / Cohen's κ / Krippendorff's α） 只需替换 mqm_re_annotate 内部循环即可，外部契约不变。

• MQM Council 2024 10 周年更新：https://www.themqm.org/
• Phrase MQM 评分：https://phrase.com/blog/mqm-quality-metric/
• Lokalise Translation scoring：https://docs.lokalise.com/en/articles/11631905-scoring-translation-quality
• Google 二次标注论文（Riley et al., 2025-10-28）

#Roadmap（未在 0.1.0 落地，0.2.0 候选）

#合规姿态：EU AI Act + GDPR（2026-08 声明）

• GDPR Art. 28 要求翻译记忆若含 PII 必须本地化处理。yimai「纯本地、零云端依赖」天然合规—— TM/术语/学习闭环全在 127.0.0.1:8787 闭环，tm_store.json 原子写持久化在本地工作目录。
• EU AI Act 2024-2026 高风险 AI 系统需可审计、可追溯。yimai 是 non-deployable unit： adopter 集成进翻译产品时承担部署者责任。引擎提供白盒路径（explain_card / activation_path / prediction_path / value_breakdown）+ MQM 严重度尺度作为 可审计凭证。
• 不引入 OAuth / ID-JAG / Mcp-Session-Id 等云端鉴权方案——保持"纯本地"定位；如需 企业版鉴权，由 fork 独立分支承担。

#MCP server 集成 Claude Code / Cursor / Gemini CLI

#For Agents / Integration

AI agent 拆箱即用：克隆后先读 AGENTS.md（项目结构、构建/运行/消费指南、Windows 前置、MoonBit 坑、多 harness 接入），再跑 scripts/dev.ps1 一键起服务——无需人工配置。

• Claude Code 用户：直接读 CLAUDE.md 即可
• Gemini CLI 用户：直接读 GEMINI.md 即可
• 其他 harness（Cursor / Cline / Continue.dev / Roo Code / Windsurf / Copilot / Codex / Devin）：用 AGENTS.md 多 harness 接入表

安装为 WorkBuddy skill（可选）：仓库内 docs/skill/SKILL.md 是编排手册（frontmatter 含 agent_created: true，触发词 + 26 端点 API 手册 + MCP 接入 + 数据契约）。复制到 ~/.workbuddy/skills/yimai-prophecy/ 并重启 WorkBuddy 后即成为可用技能。

cd yimai_prophecy_moonbit
moon test --target wasm-gc        # green ⇒ engine is usable
cd cmd/main && moon build --target wasm-gc && moon run .

{ "mcpServers": { "yimai": { "url": "http://127.0.0.1:8787/mcp" } } }

#Contributing

#License

#Acknowledgements

• Built on the MoonBit language and its core (json, math) packages.
• README structure follows conventions of high-star open-source projects (e.g. sharkdp/bat, BurntSushi/ripgrep) and the idiomatic MoonBit library style of moonbit-community/moon_elk.
• 完整架构方案（评审叙事 + 工程附录 13 章）：译脉·先知2.0_完整架构方案.md（黑客松定位、三层架构、确定性/零依赖/白盒三大卖点、S1 实证、增强路线图 S/M/L → 端点映射）。