GACEE · Global Association of Cultural and Educational Exchange

汉桥项目
落地方案HanQiao · Delivery Plan for a Three-country Pilot · v3.0

本方案旨在说明汉桥(AI 中文学习平台)在三个国家、各一所学校的落地路径,并以试点数据作为下一阶段对外合作的依据。全文按「对外洽谈 → 需求调研 → 技术验证 → 落地实施」四条工作流展开,附应用场景、技术路线、备忘录与合同参考样本。时间窗口只写月份与所需时长;所有签约方在正式签署前均为「待定」;凡属待验证的假设,均在文中单独标注。

版本 3.0 · 2026 年 9 月依据:GACEE 协会介绍 · JE 汉桥架构方案密级:协会内部 · 对外版另出
00 · Executive summary

执行摘要One page

本项目的核心任务是将远程中文老师、可编排的智能体与离线优先的学习端纳入三所学校的常规课表,在弱网、停电与共享设备条件下完成六周教学,并取得可核证的运行数据。

项目一句话

以老师为根,以三国三校为试金石,以弱网和离线为默认,把合规埋进地基里。对外入口为 learn.gacee.org,证书验真为 verify.gacee.org。

目标

9 月完成立项、候选校筛选与技术验证启动;10 月底冻结 MVP;11 月起三国各一所学校开课六周;年底前完成复盘并决定扩大范围。

试点格局

三个国家 × 各一所学校 × 每校 20–40 名学员。三国分别覆盖英语、斯瓦希里语、法语三种辅助语言与三套数据法规,用同一套系统验证「可迁移」。

落地角色

协会(内容、师资、国际资源、项目设计)+ 试点校(学员组织与现场)+ 远程中文老师(教学与签发)+ 技术平台(系统与 AI)。签约主体在各国法务确认后确定。

八条原则

老师主导、无幻觉、普惠优先、弱网离线默认、母语脚手架、合规入地基、先轻后重、对齐标准与当地体系。

四条并行工作流与时间窗口

工作流时间窗口所需时长交付负责方
A 对外洽谈9 月起持续首轮 4 周,之后滚动三国候选校短名单、试点合作备忘录与意向书(签约方待定)、老师协议、家长同意书模板协会主导,Ray 负责非洲,刘老师负责师资
B 需求调研9–10 月每国 2–3 周,三国并行需求文档 v1、五类用户画像、三国实地网络与设备普查报告、合规登记册 v1协会与平台联合,当地伙伴执行测量
C 技术验证(阶段 0)9–10 月4–5 周8 张 PoC 测试卡全部有结论,决定「直播为主」还是「录播为主」;架构决策记录 ADR-1 至 7 落笔平台
D 落地实施(阶段 1–3)9 月中起MVP 开发 6 周;试点准备 2 周;每校试点 6 周;复盘 2 周;阶段 2 约 12 周MVP 冻结 → 三校开课 → 复盘报告与单位成本 → 阶段 2 验收联合项目办公室
前置判断

在两个月内使真实班级投入使用,前提是将 MVP 界定为「录播 + 现成会议工具直播 + 可安装网页应用(PWA)离线练习 + 伴学与复习两个智能体」。原生离线安卓 App 需要 10–12 周成熟期,安排在试点开课后启动,于阶段 2 发布测试版。此安排并非缩减范围,而是先在网页端实现「弱网离线」原则,再迁移至原生端。

成功标准(试点复盘时逐条验收)

  1. 三国各一所学校、每校 20–40 名学员完成 6 周学习,出勤率(含回放)与作业提交率有系统记录。
  2. 老师每周实际投入课时下降,学员练习量上升,两者都有日志数据;每老师课时服务的学员数有实测倍数。
  3. 智能体答疑的抽样准确率与「置信不足即转人工」的比例被度量,没有一起未被发现的错误判定。
  4. 断网与停电条件下学员仍能完成练习并在联网后同步,零数据丢失;三国网络实测数据可比。
  5. 学校数据隔离、监护人同意、跨境传输三项合规要求在三国各有书面证据。
  6. 产出「每学员每月成本」实测值,成为之后与教育部门、企业与基金会谈价格的唯一依据。
01 · Landscape & roles

落地格局与角色Who does what on the ground

四位一体的长期结构

出版

沉淀和筛选优质内容:作文大赛作品集、课程教材、文化读本;反向成为平台图书馆的书源。

未来书院

形成课程与真实教育场景:中华文化、世界文明、AI 与未来、成长实践四大板块。

AI 平台(汉桥)

把内容与教学能力规模化:远程老师 + 智能体编排 + 离线学习 + 在线图书馆。

协会

连接学校、老师、专家、国际资源与项目,是四者的枢纽与品牌持有者。

汉桥在协会域名下对外提供三个入口:平台入口 learn.gacee.org、证书验真 verify.gacee.org,以及阶段 2 接入协会网站的单点登录。

落地角色与分工

角色负责拥有 / 决定不做
协会 GACEE教育内容、教师资源、国际资源、项目设计、学校与教育部门关系、品牌、对外签约品牌、课程内容版权、学员关系;作为数据控制者决定数据用途不自建技术团队
试点校(每国一所,待定)学员组织、课表时段、场地与设备、当地助教、家长沟通与同意书回收本校学员的管理权与本校数据的联合控制不承担系统建设成本;试点期免费
远程中文老师(中国侧)录课、直播、复核队列、签发关键判定、参与教研教学判断的最终决定权;音色与录播的授权范围不做技术运维,不处理学员敏感数据
当地教育主管部门(阶段 2 起接触)准入、课程体系衔接、认证路径、公办校渠道本国认证与准入试点期不介入日常运营
技术平台(服务提供方)系统、AI 能力、运维、安全与合规工程;按协会指示处理数据平台与智能体底座的技术实现不以协会名义对外;不直接向学校收费
企业客户(阶段 2 起)为本地员工开委托班,提供学员名单与 HR 对接人本企业学员的报表与证书用途不接触其他租户数据

核心角色与资源人

角色负责在本方案中的位置
项目发起人协会负责人决策、预算、对外关系、四位一体的统筹各里程碑的批准人;对外签约的授权人
非洲商务资源负责人Ray三国候选校、当地伙伴、教育部门、支付与本地渠道、非洲市场伙伴工作流 A 非洲部分;三校试点准备与运营的校方接口
课程体系、资源包与师资负责人(教研负责人)刘老师中文课程体系、内容资源包、中国侧老师的招募、培训与排班、内容终审§10 教研、老师协议、MVP 内容门禁
教育专家资源张老师教研评审、专家导学、校长与学校网络(具体分工待协会确认)教研评审与校长网络支持
出版与媒体资源李社长作品集与教材出版、媒体报道、内容审校(具体分工待协会确认)图书馆书源与「出版」一环
平台负责人(技术)平台技术路径、架构决策、平台团队调度工作流 C、D 的技术批准人

对外收入与资助来源(与协会公益属性并存)

  • 机构席位:学校、华校、培训中心按学员席位购买,自带免费基础版,完整版按席位升级。价格待定,不在本方案预设;试点期三校全部免费。
  • 企业委托班:中资出海企业为本地员工开班,按班结算,出结业证书与学习报告(阶段 2 起)。
  • 协会自有收入:活动报名费、研学费、作文大赛报名费、会员费,走协会自己的收款账户。
  • 公益资助:EduLink-Africa 与公益讲座由基金会与捐赠支持,普惠版免费给普通学校;复盘报告是申请资助的核心材料。
02 · Three-country pilot

三国试点设计Three countries, one school each

选择三个国家的目的在于以最小样本验证同一套系统在不同辅助语言、网络画像与数据法规下的可迁移性。每国仅设一所学校,以保证单个班级的试点深度与数据质量。

候选国家(建议,待协会确认;具体学校与签约方待定)

席位建议国家为什么辅助语言数据法规(需法务确认)备选
P1尼日利亚协会已有线上中文课与学校关系,最现成;西非英语区最大市场;供电与移动网络条件最能检验弱网离线英语NDPA 2023 与 NDPC 实施规则加纳
P2肯尼亚东非枢纽,斯瓦希里语脚手架的天然试验场;海底光缆与 4G 覆盖好于西非,可作为「网络较好」对照组斯瓦希里语 + 英语Data Protection Act 2019(ODPC)坦桑尼亚
P3科特迪瓦西非法语区经济中心,验证法语界面与法语讲解脚手架,为阶段 3 的法语区扩展提供依据法语2013 年个人数据保护法(ARTCI 监管)塞内加尔 · 喀麦隆

若协会希望其中一席放在非洲以外验证亚洲市场,印度尼西亚是备选(印尼语界面包已就绪);但三席同在非洲更利于 Ray 统一对接与复盘可比。

候选校四维打分(每维 1–5 分,总分 ≥ 14 进入短名单)

网络与供电

校内 Wi-Fi 或 4G 可用时段、停电频率、是否能设一个局域缓存点。

需求与组织力

已有中文课或明确需求、能固定 6 周课表、有一位愿意当助教的老师。

合作方成熟度

校长决策链短、家长沟通渠道畅通、愿意签意向书与同意书流程。

合规可行

监护人同意可落地、数据跨境有合法路径、无「不可做」结论。

每校统一规格

规格说明
学员20–40 人,按年龄与基础分 1–2 组低于 20 人样本不足,高于 40 人老师复核队列超载
时长6 周,每周:1 次直播(30 分钟)+ 1 节录播 + 每天 15 分钟练习三校开课时间错开不超过 2 周,便于同一老师排班
老师每校 1 位中国侧主讲老师,三校可共用 1 位备课与复核老师录播复用,一位老师带多校
设备学员自带手机(含共享)+ 学校 2–4 台共享设备 + 1 台缓存点设备(可选)2 GB 内存安卓机为验收基线
当地助教每校 1 人,每周 4 小时点名、设备、现场支持、执行网络测量
费用试点期全部免费;协会承担第三方直接成本设备与流量补贴按校按月记录,进入单位成本模型
产出学习报告、结业证明(协会签发,可验真)、学校周报、教育部门版学期报告证明与官方 HSK 成绩严格分开

三国对照设计

三校使用同一套内容(HSK 一级前 6 周)与同一套系统,只有三个变量不同:辅助语言、网络画像、法规。复盘时按这三个变量做横向对照,得到「哪些能力是可迁移的,哪些必须按国家定制」的结论,这是阶段 3 规模化的依据。

03 · Workstream A

对外洽谈Sequence, counterparties, materials

洽谈遵循「先内后外、先近后远」的顺序:先完成方案、试点规格与合作方要求的整理,再以成熟方案对外沟通。所有对外承诺仅来自事实卡;所有签约主体在法务确认前标注「待定」。

利益相关方地图

对象他们要什么我们要什么接触时机负责
协会理事会项目符合公益属性、可控风险、品牌加分立项决议、预算框架、对外签约授权9 月,首 1 周协会负责人
三国候选校(每国 2–3 所候选,最终各签 1 所)更稳定的中文课、学生成果可见、不增加成本、数据安全学员、场地、网络与设备共享、助教、家长同意流程9 月起,4 周内签意向书Ray + 协会负责人
中国侧远程中文老师合理课时与报酬、工具好用、不被 AI 替代、录播与音色授权清楚参与教研与试点、签发关键判定9 月起刘老师
当地助教 / 合作伙伴津贴、培训、能力认证现场执行、测量、点名意向书签署后Ray
教育专家资源(张老师)· 出版与媒体资源(李社长)项目有专业水准与影响力;可出版的成果与报道素材教研评审、专家导学;作品集与教材出版、媒体报道9 月起协会负责人
当地教育主管部门符合当地课程与认证体系、数据合规准入、认证衔接、公办校渠道试点出数据后(12 月起)Ray + 协会负责人
华校 / 国际学校 / 大学中文系优质课程与考点合作渠道与考点阶段 2协会
中资出海企业员工中文培训有报表、有证书企业委托班MVP 稳定后(12 月起)协会
基金会 / 捐赠方可衡量的公益成果普惠版资金复盘报告后协会

洽谈顺序与每一步的目标

9 月 · 第 1 周 · 内部对齐
理事会通过立项、预算框架、品牌与域名决议、对外签约授权。产出:理事会决议。
9 月 · 第 1–4 周 · 候选校筛选与意向书
Ray 按四维打分出三国短名单,每国走访或视频访谈 2–3 所,签 1 所意向书(LOI);同步启动家长同意书本地化。产出:LOI ×3、同意书三语模板。
9 月 · 第 2 周起 · 老师合作协议
刘老师牵头,与首批 2–3 位中国侧老师签课时协议,含录播版权、音色复刻授权(单独签)、数据保密。产出:教师协议 ×3、音色授权书。
10 月 · 试点合作备忘录
与三所学校(及其所属教育机构,若有)签试点合作备忘录(MOU)+ 数据保护附件;样本见 §09。产出:MOU ×3、DPA 附件 ×3。
12 月起 · 教育部门与渠道
带着试点数据(出勤、作业、准确率、成本、三国对照)去谈教育主管部门与华校。产出:合作框架协议(草案)。
12 月起 · 企业委托班
向 2–3 家中资企业提案,先做 1 个班的付费试点。产出:委托班服务协议(样本见 §09)。

洽谈材料包

  • 协会介绍:精简为 15 页版本,用品牌手册统一视觉。
  • 汉桥一页纸:定位、八条原则、三国三校试点规格、对合作方的要求与回报。
  • 试点校版说明(英 / 法 / 斯瓦希里语):学生每周做什么、需要学校提供什么、数据怎么保护、免费期多久、结束后有什么。
  • 教育部门版说明:与 HSK 标准对齐、与当地课程体系衔接的路径、数据主权承诺、教育部门版学期报告样例。
  • 企业版说明:委托班流程、HR 报表样例、证书样例、按班报价(待定)。
  • 演示环境:learn.gacee.org 的角色演示账号(学员、老师、学校、企业、运营),随时可展示;试点开课后换成真实班级的脱敏视图。
  • 事实卡:所有对外数字的唯一来源,含出处与更新日期。
洽谈纪律

不做口头承诺指标(时延、准确率、覆盖国家数);所有对外数字来自事实卡;涉及未成年人与数据的问题一律回答「按当地法规设计,法务确认后书面回复」;不以任何一方名义替另一方承诺。

04 · Workstream B

需求调研Ask, measure, observe — in three countries

调研旨在查明三所试点校中老师与学员的实际学习条件:可用设备、每周可用时段、学习方式与网络环境。调研结论决定 MVP 范围。三国并行,每国 2–3 周,使用同一套工具包,结果可横向对照。

调研对象与样本(每国)

对象样本方法要回答的问题
试点校校长与教务2 人结构化访谈 60 分钟(附录 A)课表时段、场地、设备数量、供电情况、家长沟通方式、收费习惯、认证诉求
当地带班老师 / 助教2 人访谈 + 一节课观察课堂组织方式、学生水平分布、母语与英语 / 法语 / 斯瓦希里语使用、对 AI 的顾虑
中国侧远程中文老师3 人(三国共用)访谈 + 一周工作日志备课与批改时间分布、跨境上课的网络体验、最希望被分担的工作
学员15 人,分年龄段问卷(当地语言)+ 6 人小组访谈手机拥有与共享情况、每天可学习时长与时段、流量预算、学中文的动机
家长10 人问卷 + 电话访谈付费意愿与能力、对孩子数据的态度、期望的成果形式(证书 / 成绩 / 能对话)
现有中文课程协会现有课程与学校现有教材内容审计已有课件、进度、考核方式、可复用的内容资产

实地测量工具包(由当地助教按步骤执行,三国同一模板)

网络测量

  • 在学校 3 个时段(上午、下午、晚间)各测一次下行 / 上行 / 时延 / 丢包,记录运营商与制式(2G / 3G / 4G / 5G)。
  • 做一次 20 分钟 Zoom 音频通话到新加坡,记录掉线次数与主观清晰度(1–5 分)。
  • 下载一个 50 MB 测试文件,记录耗时与是否中断;分别测 Cloudflare、阿里云 CDN 与直连对象存储。
  • 记录一周内停电时段与学校备用电源情况。

设备与支付普查

  • 学员与学校设备:品牌型号、安卓版本、内存、剩余存储、是否多人共享。
  • 学校是否有可共享的平板 / 电脑 / 投影 / 音箱;能否放一台常开设备作为局域缓存点。
  • 每月流量套餐费用与用量;常用支付方式(移动钱包 / 现金 / 银行转账)。
  • 学校本地语言与授课语言;学生对英语 / 法语 / 斯瓦希里语的阅读水平。

产出物

产出内容时限
用户画像 ×5远程老师、当地助教、学员(儿童 / 青少年 / 成人)、家长、校方管理者;每国标注差异调研第 2 周末
用户旅程图 ×3一周学习旅程(有网 / 弱网 / 断网三条线)、老师一周工作旅程、学校开班到结业旅程调研第 2 周末
三国实地测量报告网络、设备、供电、支付的实测数据,带原始记录;三国对照表调研第 3 周末
产品需求文档 v1功能需求(MoSCoW 分级)、非功能需求(弱网、离线、低端设备、多账号、辅助语言)、数据与合规需求调研第 3 周末
合规登记册 v1逐国法规、义务、责任人、证据(见 §11)调研第 3 周末

调研与技术验证并行:调研第一周的网络测量结果直接喂给 PoC-1 与 PoC-2,不等报告写完。

05 · Workstream C

技术验证与阶段 0Prove the risky parts first

技术架构、Agent Harness、JE 引擎编排与 TIDAR KAG 见 §06;本章说明验证方法与通过标准。八张测试卡全部有结论,才进入 MVP 冻结。三国的网络类测试卡各跑一遍,其余测试卡跑一遍即可。

复用与新建(摘要)

直接复用(已上线、已验证)

  • Go 后端、Postgres、新加坡 Kubernetes、多租户隔离与行级安全
  • JE Engine / JE Agent 单回合工具循环、TIDAR KAG 有出处答疑
  • 知识图谱与掌握度(compute-on-read)、Wilson 下界判定
  • 多智能体批改 + 复核、老师音色复刻(CosyVoice)、拍照识别(Qwen-VL)
  • 邮箱验证码、线下收款登记、可验真证书、站内助手与工单、七语界面包

新建(本项目独有)

  • 跨境弱网直播接入与三级降级(先用现成会议工具深链)
  • 录播转码、多码率分发、离线内容包与学校缓存点
  • 离线优先客户端:PWA 先行,原生安卓后置;共享设备多账号
  • 每用户智能体花名册与四层记忆(档案 / 掌握度 / 情节 / 工作)
  • 中文语音识别与发音评测管线(阶段 2);HSK 课程图谱子树与题库(教研工程)

八张 PoC 测试卡(9–10 月,4–5 周)

#验证什么方法通过标准不通过则
PoC-1跨境直播可用性(三国各测)中国老师经 Zoom 与腾讯会议各上 3 节 30 分钟课到每所试点校,覆盖早中晚三个时段;5 名听众打分,记录掉线与降级触发音频主观清晰度 ≥ 3.5 / 5;每节掉线 ≤ 1 次且 30 秒内恢复;三级降级实际触发过并成功该国转「录播为主 + 每周一次音频答疑」
PoC-2内容分发(三国各测)从校内 4G 下载 50 MB 课程包,比较 Cloudflare、阿里云 CDN、直连 OSS 各 5 次中位耗时 ≤ 8 分钟,断点续传成功率 100%改夜间预下载 + 学校局域缓存点
PoC-3低端安卓可运行2 GB 内存、安卓 9 的机器上安装 PWA,做 20 分钟练习冷启动 ≤ 5 秒,练习无崩溃,占用存储 ≤ 150 MB,首包 ≤ 1.5 MB(gzip)削减动画与字体包,改纯文本模式
PoC-4离线练习与同步飞行模式完成 3 组练习与 1 次模拟测验,24 小时后联网数据零丢失,同步 ≤ 60 秒,冲突按「服务器权威 + 事件合并」正确处理重做同步引擎,MVP 缩到只离线复习卡
PoC-5伴学智能体记忆闭环1 名学员 5 次会话,验证档案、掌握度、情节记忆的写入与召回;50 道课程内问题评测有出处率 100%,错误率 ≤ 2% 且全部触发转人工,第 5 次会话能正确引用前 4 次的薄弱点限制智能体只做复习与提醒,答疑全转老师
PoC-6多账号共享设备一台手机 3 个学员账号切换,各自离线数据隔离切换 ≤ 3 秒,数据不串,PIN 保护,本地各自加密MVP 改一机一号,共享作为阶段 2
PoC-7本地收款登记与试点校确认未来收费方式;测一笔移动钱包或学校代收登记与对账一笔款项能在系统内被记录并对账试点期全部免费,支付放阶段 2
PoC-8合规可行(三国)法务出具三国数据法 + 新加坡 PDPA + 中国 PIPL 适用性备忘与三语家长同意书备忘签字;DPA 附件完成;无「不可做」结论调整数据流(如老师数据不出境)后再进 MVP

通过标准中的数字是阶段 0 的工作阈值,不是对外承诺;PoC 结束后按实测修订。

MVP 技术范围(10 月底冻结)

模块MVP(10 月底)阶段 2(次年 1–3 月)
直播Zoom 或腾讯会议深链,平台管排课、点名、录制回收、按地区推荐入口商用 RTC 适配层(按 PoC-1 结论)
录播老师录课上传,多码率 HLS,字幕(中 / 英),课程包下载老师音色旁白、法语 / 斯瓦希里语字幕
客户端可安装网页应用:离线练习、复习卡、录播观看、多账号切换原生安卓 App(测试版)、iOS
智能体伴学 + 复习两个智能体,档案与掌握度记忆,答疑有出处、置信不足转老师发音教练、情景对话、评测、班主任智能体
评测客观题自动判分、拍照交作业识别 + 批改 + 老师复核口语评测、模拟考体系、可验真结业证明
老师端班级学情、待复核队列、预警列表、AI 提议收件箱(网页)班主任智能体草拟沟通与排课
学校端 / 企业端名册导入、出勤周报、设备与同意书看板教育部门版学期报告、企业 HR 看板、批量证书验真
运营端复用超管台:租户、学校、班级、学员、内容审核、语言包管理、客服工单渠道与合作方管理、结算、AgentOps 对外看板
辅助语言界面中 / 英 / 斯瓦希里语 / 法语;答疑英文脚手架法语与斯瓦希里语讲解脚手架、阿拉伯语界面
在线图书馆作品集与未来书院读本的网页阅读器:书架、进度、划线笔记、生词点读老师音色朗读、共读批注、出版新书自动上架、阅读数据回流知识图谱

架构决策记录(阶段 0 内落笔)

  • ADR-1 租户模型:GACEE 为一级租户,每个国家为区域标签,每所学校为子租户;数据按学校隔离,跨校汇总只在协会视图;Postgres 行级安全 + 自动化隔离测试。
  • ADR-2 直播来源:MVP 用现成会议工具深链,媒体接口做适配层;切换商用 RTC 不改业务代码。
  • ADR-3 离线策略:PWA 先行,Service Worker + IndexedDB 存内容包与练习;原生 App 复用同一同步协议与内容包格式。
  • ADR-4 模型路由:所有生成、识别、合成走 Model Gateway;中文教学答疑默认 Qwen 系,复核可用 Claude;断网端侧不跑模型。
  • ADR-5 记忆存储:档案与掌握度结构化入库;情节记忆只追加、可导出、可删除;记忆不当事实。
  • ADR-6 数据驻留:核心数据存新加坡;老师个人数据在中国侧按 PIPL 最小化;学员数据出境依各国法规与监护人同意;不在中国侧落地媒体节点。
  • ADR-7 成本护栏:每租户每月模型调用与流量预算,超限降级到录播与本地复习,避免费用失控。
06 · Architecture & technical route

技术架构与技术路线JE Engine, the Agent Harness, and TIDAR KAG

汉桥的技术基座为 JurisEdu 已投入生产运行的 JE 引擎、Agent Harness 与 TIDAR KAG。本项目在该基座之上新增三层能力:跨境弱网媒体层(直播 / 录播 / 内容包)、离线优先客户端(PWA / App)以及按学员个性化记忆的智能体编排层。本章说明各层的组件规格、接口约定与验收指标;凡涉及时延、带宽、准确率的数值,均为需在阶段 0 实测的工作假设。

架构原则

云中立(Kubernetes 声明式部署、S3 兼容存储、OpenTelemetry 观测、Model Gateway 抽象模型供应商);数据驻留新加坡;智能体只提议不执行写操作;所有学习产物挂追加式哈希链证据;任一网络段退化均有确定的降级路径。

一、地理拓扑:三段网络路径

中国 · 老师接入段

  • 老师端(Web / 桌面):直播推流、板书批注、复核队列、学情看板
  • 跨境接入:会议工具的中国大陆节点或全球加速;平台仅传信令与业务数据,不在中国侧部署媒体节点(ADR-6)
  • 班主任智能体(阶段 2):排课、预警、沟通草拟

核心区域 · Agent Harness(新加坡)

  • 客户端网关:Ingress、OIDC 认证、按租户限流、媒体信令
  • 媒体层:会议工具适配 / RTC 适配、录制回收、转码、VOD 源站
  • 领域服务:课程、排课、评测、学员、证书、收款登记
  • JE 引擎 · JE Agent 混合编排:Router → 专才 Agent → 工具 → Verifier → 转人工
  • 知识与记忆:中文课程图谱、掌握度、四层记忆
  • 数据 / 存储、观测、安全、合规、审计留痕

海外 · 学员接入段(三国试点校)

  • 边缘 CDN / 学校缓存点:录播与内容包就近分发
  • 学员端 PWA(离线优先):本地库、内容包、端侧练习与判分
  • 专属智能体花名册:伴学、复习先上;发音、对话后置
  • 共享设备多账号:每账号独立加密本地存储

实时与 API 流量走加速链路;录播与内容分发走 CDN;三段路径各自优化,任一段退化均触发本章第六节的降级路径。

二、Agent Harness 十层:智能体的构建、运行、评估与治理底座

Harness 将智能体的全生命周期能力做成十层可复用底座,汉桥为其上的一个垂直实例;后续其他语种或学科项目复用同一运行时。新增一个智能体只需提交「定义清单 + 工具绑定 + 评估集」,不涉及基础设施改动。

组件规格汉桥用到什么
用户层学员 · 老师 · 运营 / 教务 · 学校 · 企业 · 当地助教;按租户与学校隔离六类角色,每类一套界面(§07)
入口层PWA / App · Web · 直播 / 录播端 · REST API · Webhook · 工作流触发;OIDC 单点登录MVP 用 PWA + 会议工具深链;学校端名册导入 API
编排层JE Agent 混合编排:确定性 Router → 专才 Agent → Verifier 复核 → 转人工;任务拆解、工具选择、记忆读写、反思修正、多 Agent 协作、状态与上下文管理;智能体定义清单(YAML:角色、工具白名单、记忆作用域、评估集、护栏策略)伴学 + 复习先上;发音 / 对话 / 评测 / 班主任按阶段接入
能力层内置工具:图谱检索 · ASR · 发音评分(GOP) · TTS · 题库 · 间隔重复(FSRS) · 批改 · 手写识别;外部工具:对象存储 · CDN · 会议工具 / RTC · 消息 · 收款登记;工具注册表区分只读 / 写工具复用多智能体批改、Qwen-VL 识别、CosyVoice 音色;新建 ASR / GOP
模型层Model Gateway:按槽位(生成 / 复核 / 识别 / 合成)路由;成本、时延、驻留感知选路;熔断与故障转移;版本灰度;响应净化。Model Zoo:JE 引擎 · Qwen / Model Studio · Claude · Gemini · 开源 · 私有网关已上线(本章第三节 R4)
数据层中文课程知识图谱 · 向量库 · 结构化业务库(Postgres,行级安全) · 缓存(Redis) · 对象存储(S3 兼容) · 个性化记忆;治理:质量、血缘、访问控制、审计、脱敏、生命周期课程图谱扩 HSK 体系(本章第四节)
评估监控智能体评估(自动化 + 人工 + 业务指标) · 全链路 Trace / Span(OpenTelemetry) · Token 与成本 · 有出处率 · 错误率 · 护栏命中 · 转人工率 · 告警看板试点 KPI 直接取自此层
治理安全身份权限(RBAC + 租户作用域) · 策略管理 · 内容安全 · 合规审计 · 隐私脱敏 / 加密(静态 AES-256,传输 TLS 1.3) · 成本与配额租户预算护栏(ADR-7)
基础设施Kubernetes · 容器 · 服务网格 · 弹性伸缩;CI / CD 与不可变镜像;多活、备份容灾、监控告警;跨境加速、多 CDN沿用新加坡集群与发布范式
智能体生命周期定义 → 构建 → 测试评估 → 灰度部署 → 监控 → 优化;发布门禁 = 评估集三项指标不退步每个智能体有评估集与灰度通道

三、JE 引擎与混合编排:已在生产环境验证的六项能力

以下六项能力已在 JurisEdu 生产环境上线并经端到端验证,汉桥直接继承。每一项均以「有依据、可审计、人做主」为约束。

#能力实现方式对汉桥的意义
R1意图 / 领域路由确定性启发式路由(无模型调用、零延迟)将请求分到学习 / 账户 / 日程等领域,先注入该领域的 grounded 上下文再作答:route → ground → answer学员提问先被分类,学习类问题自动附带其校准后的薄弱知识点
R2跨能力工具注册表只读工具接入教练的工具循环;写工具使用原生 function-calling,且一律 propose → approve → execute;审批记录以 SHA-256 链式哈希(前一记录哈希 + 本记录内容)写入追加式账本「智能体建议练习、老师批准生成」即老师签发关键判定的技术形态
R3持久学习者记忆跨会话记忆仅有两条写路径:由真实掌握度快照派生的观察,以及用户逐字自述的目标;重复出现累加显著度;证据不足标「样本尚少」按学员积累的记忆在结构上不可能编造
R4模型网关叠在 PII 去标识 → 新加坡驻留 → 故障转移 → 用量计量链之前,只增加槽位策略与响应净化;网关错误自动回退直连按成本与合规选路,且不绕过任何护栏
R5单次运行轨迹每回合记录 route / ground / recall / tool / verify / end 步骤,只存安全元数据,不存问题与答案原文试点期「为什么这样答」可回放;转人工比例可度量
R6教育 TIDAR KAG本章第四节低幻觉答疑与掌握度判定的核心
创作链

每件学习产物(作业、作文、评测、证书)都有一条追加式哈希链证据。模型没有写链的工具;写失败必上浮为错误而不是静默忽略。创作链是汉桥证书与学习报告可信的技术来源。

编排回路(单回合)

步骤动作依据时延预算(假设)
1 输入学员的题、问或语音;会话开始注入档案记忆与掌握度快照(as-of 当前时间)R3
2 Router确定性路由定领域与意图;召回相关情节记忆(Top-k,按显著度与时间衰减)R1 + R3≤ 50 ms
3 专才 Agent伴学 / 复习 / 发音 / 对话 / 评测 / 班主任按定义清单调用白名单工具;答疑走 TIDAR KAG 检索并带出处R2 工具注册表≤ 2.5 s(P50)
4 Verifier一致性检查;批改由评分与复核两个独立 Agent 一致才写入,不一致取复核结论并降级为「待老师」多智能体批改范式≤ 1.5 s
5 回答或转人工有出处的个性化回答;置信不足、主观题、写动作一律进入老师队列弃权门 + propose → approve
6 写记忆会话摘要写情节记忆;作答事件更新掌握度;轨迹入库R3 + R5异步

四、TIDAR KAG:低幻觉的来源

TIDAR KAG 为平台自研的「知识图谱可溯源生成」,区别于通用向量 RAG。先在法律域验证,再移植到教育课程图谱并已上线。三个支柱构成其不编造的技术保证。

时序 Temporal

任何判断均基于某一时间点之前的真实作答快照(as-of);两个时间点之间可做逐知识点差分:进步、回落、新评估、无变化,如实呈现。

可溯源 Provenance

每个结论挂真实证据:作答记录、错题干、课程图谱中的先修 / 相关邻域;回答带出处,无出处则不答。

可校准 Calibrated

掌握等级不用点估计,而用 Wilson 置信区间下界(z = 1.96)判定;作答少于 3 题一律「证据不足」弃权,0 题为「尚未评估」。

课程图谱结构

层级
体系 → 学段 → 学科 → 领域 → 主题 → 知识点;只写权威课标明列的结构
汉桥子树
「国际中文 HSK 3.0(三等九级)/ YCT」体系:词汇、语法点、汉字、场景四类知识点,每个节点挂等级、来源、示例与先修边
边类型
先修(prerequisite)、相关(related)、场景包含(scene-contains)、易混(confusable)
证据节点
作答事件(题目 ID、结果、时间、来源端)、批改结论(双 Agent 一致)、老师签发记录
掌握度
compute-on-read:按知识点聚合 as-of 快照,输出 Wilson 下界与样本数;不落地缓存结论,避免过时
低幻觉机制规则汉桥中的落点
知识边界答疑只在课程图谱与教材范围内检索生成;超出范围回复「本课程范围外,已转老师」HSK 一级子树 + 6 周教材切片
弃权门置信不足或题型主观即弃权,交人处理弃权直接进入老师复核队列,周报统计转人工率
记忆非事实记忆只作检索证据与个性化风格,事实以图谱与实时掌握度为准学员换班、重学时自动纠正
双 Agent 一致评分与复核两个 Agent 独立判断;不一致取复核并降级为未知;只有明确对错才写入图谱拍照作业与客观题评测沿用;口语评分分歧项转老师
写权分离智能体只提议;人批准后由系统执行并写哈希链安排练习、调整进度、发证书均走审批
评估集每次发布跑课程内评估集:有出处率、错误率、转人工率三项不退步阶段 0 建 50 题,MVP 扩到 200 题,阶段 2 扩到 500 题

五、每用户智能体花名册与四层记忆

伴学 Agent

规划路径、动态调难度、串讲知识点;答疑走 TIDAR KAG 带出处;离线时按预打包任务运行。

复习 Agent

按 FSRS 遗忘曲线与掌握度编排间隔重复卡片;调度逻辑在端侧运行,离线可用。

发音教练 Agent(阶段 2)

ASR + 声调 / 音素评分(GOP),逐字纠音,用老师音色示范正确读法。

情景对话 Agent(阶段 2)

点餐、问路、面试等角色扮演,老师音色 TTS,纠错温和。

评测 Agent(阶段 2)

组卷、模拟考、多智能体批改与复核;口语题接发音管线;主观题弃权交老师。

班主任 Agent(老师侧,阶段 2)

排课、进度与风险预警、缺勤补课建议、家校沟通草拟;所有动作走提议审批。

记忆层内容存储读写时机数据主体权利
档案记忆目标等级、母语 / 辅助语、时区、设备与网络画像、纠错偏好、可用时段结构化(Postgres)会话开始注入;设置变更时更新可查看、可修改
语义 / 掌握度每个知识点、字词、语法的掌握度与易错点知识图谱(compute-on-read)每次作答与评测后以事件更新可导出
情节记忆历次互动摘要、错因、里程碑、动机信号只追加的事件日志 + 向量库会话结束写入;下次按相关性召回可导出、可删除
工作记忆当前会话上下文会话内(Redis,TTL)会话内有效,结束即蒸馏为情节记忆不持久化

六、跨境弱网直播、录播分发与离线客户端

直播选型要点取舍
A 现成会议工具Zoom / 腾讯会议承载直播;平台通过深链集成,负责排课、点名、录制回收、课堂事件回流;按学员所在地区推荐入口成本最低、落地最快;课堂定制受限。MVP 采用
B 商用 RTC SDK需自定义课堂 / 白板 / 深度嵌入时,采用具全球实时网络的 SDK(阿里云 RTC / Agora / 腾讯 RTC / ZEGO)体验可控;按分钟计费。阶段 2 按 PoC-1 结论评估
C 自建 SFUWebRTC SFU(LiveKit / mediasoup)+ 自建 TURN + 跨境加速边际成本低但运维重;规模化后评估

三级降级与录制兜底

  • 一级:全互动直播(音视频 + 白板 + 实时问答)
  • 二级:关闭摄像头,保留音频 + 板书 / 幻灯 + 字幕 + 文字互动
  • 三级:网络不可用时转「课后 15 分钟内可回看的录播 + 智能体陪练补齐」
  • 每堂直播默认云端录制转 VOD;掉线学员计入「回放出勤」;降级事件回流到课堂事件日志

弱网工程清单

  • 自适应码率 + Simulcast / SVC,分辨率帧率自动降级,优先保音频与白板
  • Opus + DTX + FEC / NACK / PLC 抗丢包;自适应抖动缓冲
  • TURN 中继兜底 NAT 穿透失败
  • 音频优先模式:只留语音 + 板书 + 字幕

录播与内容分发

  • 老师录课或直播录制 → 源站对象存储;自动字幕(中 / 英 / 法),阶段 2 加老师音色旁白
  • FFmpeg 多码率 HLS 阶梯:240p 300 kbps · 360p 600 kbps · 480p 1 Mbps · 720p 2 Mbps · 纯音频 48 kbps;6 秒切片;AES-128 分包密钥
  • 源站 → 多 CDN,非洲边缘就近取流,断点续传;学校缓存点为常开设备上的反向缓存
  • 内容包:按节 / 单元打包(清单 + 版本号 + 校验和),可限 Wi-Fi 下载,退课可回收

离线优先客户端(PWA)

  • 本地可用:已下载录播、随堂练、复习卡、笔顺、客观题作答与判分、录音留存
  • 本地库(IndexedDB),内容包版本化;出站事件队列,幂等键;失败重试与指数退避;断点续传
  • 同步协议:事件溯源,服务器单调序号为权威;学习进度「服务器权威 + 事件合并」,笔记 last-write-wins
  • 存储预算 ≤ 150 MB,分级清理;省流模式;共享设备多账号,每账号以 PIN 派生密钥(Argon2id)加密本地库

七、评测管线、模型层与云中立矩阵

听 · 读 · 写 · 汉字书写

客观题题库自动判分;拍照作业经 Qwen-VL 识别 → 双 Agent 批改逐点点评 → 老师复核签发;手写识别校验笔顺与间架结构;每次作答以事件更新知识图谱。

口语与发音(阶段 2,需选型验证)

中文 ASR(Paraformer / SenseVoice 类)转写;强制对齐(CTC 对齐)+ GOP 发音优度打分;声调用基频轮廓(pYIN)分类;纠音示范用老师音色 TTS;大题分歧项转老师。

依赖主选(当前基线)备选可移植抽象
计算 / 编排阿里云 ACK(Kubernetes,新加坡)AWS EKS · GCP GKE · Azure AKS · 自建Kubernetes 声明式清单
对象存储阿里云 OSSAWS S3 · GCS · MinIOS3 兼容 API
数据库 / 缓存 / 向量Postgres · Redis · 向量库各云托管版或自建开源组件,无专有依赖
实时直播 / RTC现成会议工具Agora · 腾讯 RTC · ZEGO · 自建 SFU媒体适配层
内容分发 CDN多 CDN 编排Cloudflare · Akamai · CloudFront · 阿里云 CDN按目标国实测择优
跨境加速云厂商全球加速他厂加速 / 专线接入层可切换
LLM / ASR / TTSModel Studio(Qwen · Paraformer · CosyVoice)Claude · Gemini · 开源 · 私有Model Gateway
观测 / 安全日志 · 指标 · 追踪 · WAF · 密钥管理OpenTelemetry + 开源栈OTel 标准

访问与内容安全

令牌鉴权、全链路 TLS、租户隔离、最小权限;录播加密、离线包加密与回收、可选水印。

评测诚信

正式考随机题库、时限、切屏与多设备检测、口语活体读题;高利害考试建议线下考点。

Agent 护栏

越权与诱导防护、输出复核、记忆写入校验、审计留痕,避免记忆污染;媒体与 CDN 多活降级。

八、非功能指标(试点验收基线,实测后修订)

维度指标目标(假设)度量方式
低端设备冷启动 · 首包 · 存储占用≤ 5 s · ≤ 1.5 MB gzip · ≤ 150 MB2 GB 内存安卓机基线
离线同步成功率 · 数据丢失≥ 99.5% · 0客户端遥测 + 服务器对账
智能体有出处率 · 错误率 · 转人工率100% · ≤ 2% · 按实测记录评估集 + 每周抽样 50 条人工评审
直播音频清晰度 · 降级触发 · 回放可用时间≥ 3.5 / 5 · 记录 · ≤ 15 分钟PoC-1 方法沿用到试点
可用性核心区域月可用性试点期记录,不承诺;正式期 99.5%观测层
安全租户隔离测试 · 渗透测试每次发布通过 · MVP 冻结前一次CI 门禁 + 外部测试
成本每学员每月模型 + 流量 + 会议工具实测建模,按租户预算护栏月度成本表

九、技术路线图(与本方案里程碑对齐)

阶段技术目标Harness 增量时间窗口
阶段 0 · PoC三国跨境弱网实测、CDN 覆盖、低端机、离线同步、记忆闭环、多账号、收款登记、合规租户模型(ADR-1)、媒体适配层(ADR-2)、伴学智能体评估集 50 题9–10 月,4–5 周
阶段 1 · MVP录播 + 会议工具直播 + PWA 离线 + 伴学与复习智能体 + 客观题与拍照作业 + 老师复核队列 + 学校端周报HSK 一级课程图谱子树、档案与掌握度记忆、租户预算护栏(ADR-7)9 月中–10 月底,6 周开发;11 月起试点
阶段 2 · 增强ASR / GOP 发音评测、情景对话、评测与班主任智能体、可验真结业证明、原生安卓 App、法语 / 斯瓦希里语脚手架、企业委托班情节记忆与向量召回、写工具审批扩到排课与发证、评估集 200–500 题次年 1–3 月,约 12 周
阶段 3 · 规模化多国多校、商用 RTC 或混合自建、多 CDN 成本优化、教育部门报告与认证衔接多活部署、AgentOps 看板对外、单位经济模型驱动的网关选路次年第二季度起

关键待验证假设

#假设影响何时验证
1中国老师到三个目标国的跨境实时质量(时延、丢包、时段)决定各国直播为主还是录播为主阶段 0
2商用 RTC 与自建 SFU 的成本、合规与锁定成本结构与工程负担阶段 0 → 1
3CDN 在三个目标国的真实节点覆盖与单价录播体验与流量成本阶段 0
4中国侧是否落地媒体或接入节点及其合规跨境路径合法性决策前置
5中文 ASR 与发音评测模型的准确率口语评测可信度阶段 1 → 2
6端侧发音打分在低端安卓的可行性离线口语体验范围阶段 1
7教研内容产能(课程、题库、HSK 对齐)可教的级别与进度贯穿
8正式认证与发证的对接路径学习成果公信力阶段 2
07 · Applications & scenarios

应用与场景Surfaces by role, journeys by situation

本章说明六类角色各自使用的界面,以及试点期最常发生的十二个场景。每个场景列出触发条件、参与者、系统步骤、涉及组件、成功标准与异常处理,作为需求文档、测试用例与培训材料的共同依据。演示环境 learn.gacee.org 已按本章实现全部界面的可交互原型。

一、按角色的应用界面

角色入口核心界面MVP 范围阶段 2
学员PWA(手机 / 共享设备)· 学员 ID 或手机号 + PIN今日任务、课程与录播、练习与复习卡、作业拍照、直播入口、图书馆、知识星图、我的智能体、证书全部(发音与对话除外)发音教练、情景对话、原生 App
远程老师Web(桌面)· 邮箱验证码班级学情、待复核队列、AI 提议收件箱、直播与录课、排课、班级图书馆、周报全部(班主任智能体除外)班主任智能体、音色旁白
学校(校长 / 教务 / 助教)Web 或手机浏览器 · 邮箱验证码名册导入与建班、学员与设备看板、同意书覆盖率、出勤周报、学期报告名册、看板、周报教育部门版学期报告、批量证书验真
企业(HR)Web · 邮箱验证码委托班学员组、出勤与进度报表、结业证书、账单全部
运营(协会教务 / 平台)超管台 · 双因素租户与学校、内容与出版审核、AgentOps、合规登记、成本与预算、网络与分发、语言包管理、客服工单全部渠道与结算
公众 / 验证方verify.gacee.org证书编号或扫码验真,仅显示验证所需最少信息全部

二、十二个落地场景

S1 · 学员的一周(有网 / 弱网 / 断网三条线)

触发
周一录播与练习包发布;伴学智能体按学员档案(可用时段、停电预测)生成本周任务
参与者
学员、伴学与复习智能体、老师(被动)
步骤
有网:观看录播 → 随堂练 → 复习卡;弱网:音频优先模式 + 文本练习;断网:本地内容包 + 复习卡 + 离线判分,联网后出站队列同步
组件
内容包、PWA 本地库、复习 Agent(端侧 FSRS)、同步协议、掌握度事件
成功标准
三条线均可完成每日 15 分钟任务;同步零丢失;掌握度周增量可见
异常
存储不足 → 分级清理提示;内容包校验失败 → 重新下载对应切片

S2 · 一堂跨境直播课(老师在中国,学员在三国)

触发
排课到点;平台按学员所在地区推荐 Zoom 或腾讯会议入口并发出三语提醒
参与者
主讲老师、当地助教、学员、运营值班
步骤
老师从老师端一键开课 → 平台创建会议并回收会议 ID → 学员深链进入并记录进入时间 → 课中网络质量触发一级 / 二级降级 → 下课后录制自动回收转码 → 15 分钟内回放可看 → 缺席学员计回放出勤
组件
会议工具适配层、点名与课堂事件日志、录制回收、转码管线、通知
成功标准
音频清晰度 ≥ 3.5;降级事件有记录;回放在 15 分钟内可用;出勤(含回放)自动统计
异常
会议工具不可用 → 切换另一工具的备用链接;老师端掉线 → 助教主持 + 录播替代(三级降级)

S3 · 停电夜的离线练习与次日同步

触发
档案记忆中的停电预测窗口(如 19:00–21:00);伴学智能体提前将当日任务打包为离线任务,提醒移至 17:30
步骤
学员在断网状态完成复习卡与录播观看 → 事件写入出站队列 → 次日联网后按序号同步 → 掌握度更新 → 老师端看到「离线完成」标记
成功标准
PoC-4 标准;老师端可区分在线与离线完成
异常
设备时钟漂移 → 以服务器序号为准;重复提交 → 幂等键去重

S4 · 拍照交作业到老师签发

步骤
学员拍照上传(离线先存本地)→ Qwen-VL 识别手写与版面 → 评分 Agent 与复核 Agent 独立批改 → 一致项写入掌握度;不一致或主观项进入老师复核队列 → 老师逐项确认或修改 → 签发 → 结果与点评回到学员端;全过程写创作链
成功标准
客观项双 Agent 一致率有记录;老师只看分歧与主观项;零未发现错误
异常
照片模糊 → 提示重拍并给出示例;识别置信低 → 整份转老师

S5 · 智能体答疑与转人工

步骤
学员提问 → 确定性路由判为学习类 → 注入其薄弱知识点 → TIDAR KAG 在课程图谱内检索 → 有出处则作答并附来源;置信不足或超出范围则弃权,问题进入老师队列,老师在周三直播或文字回复;轨迹入库
成功标准
有出处率 100%;转人工问题有闭环记录;每周抽样 50 条人工评审

S6 · 共享设备上的三名学员

步骤
设备首页显示「谁在学习」→ 选择头像输入 PIN → 解锁该账号的加密本地库 → 学习 → 退出即锁定;各账号内容包按引用计数共享,练习数据互相隔离
成功标准
切换 ≤ 3 秒;数据不串;忘记 PIN 走学校或家长确认后重置

S7 · 老师的一周

步骤
周一发布录播与练习包(复用往期录播)→ 周三 30 分钟直播 → 每日查看预警列表(缺勤、掌握度回落、停电影响)→ 周五清空复核队列(只看分歧与主观项)→ AI 提议收件箱审批(安排练习、调整进度、补课建议)→ 周报自动生成,老师签发
成功标准
每班每周老师投入 ≤ 6.5 小时;一位老师可带三校;所有写动作有审批记录

S8 · 学校端:从名册到学期报告

步骤
教务上传名册(CSV 或照片)→ 智能体提议分班(不编造姓名)→ 学校批准 → 账号与 PIN 发放 → 同意书回收看板 → 设备与缓存点状态 → 每周出勤(含回放)与练习周报 → 学期末生成教育部门版报告(英 / 法)
成功标准
同意书覆盖率 100% 才开课;周报自动送达校长与助教;报告可导出

S9 · 企业委托班(阶段 2)

步骤
HR 提交学员名单与目标(职场场景)→ 开班 → HR 看板只见出勤、进度与结业状态,不见作答明细 → 结业证书批量签发 → 按班结算并出账单
成功标准
员工与学校学员享有同等数据主体权利;HR 权限限于报表字段

S10 · 图书馆:从作文大赛到共读

步骤
获奖作文 → 出版编辑 → 等级标注 → 自动上架图书馆 → 老师布置共读 → 学员划线、点读生词、批注 → 生词进入复习卡 → 阅读数据回流知识图谱 → 优秀课堂作品可提名下一辑作品集
组件
出版工作流、阅读器(离线整本缓存)、生词点读、共读批注、图谱事件
成功标准
每周阅读时长与生词转化率可度量;出版新书自动上架

S11 · 证书签发与验真(阶段 2)

步骤
课程结束 → 系统提议签发(出勤 ≥ 80% 且模拟考通过)→ 老师签发(写创作链)→ 协会盖章编号 → 学员获得 PDF 与二维码 → 验证方在 verify.gacee.org 输入编号,仅见有效性、姓名首字母、课程与等级
成功标准
验真响应 ≤ 2 秒;撤销即时生效;批量验真支持 CSV

S12 · 客服与工单(多渠道,智能体协助)

步骤
WhatsApp / 站内助手 / 邮件 / 学校群 / 电话进入统一工单 → 分类(下载、账号、报表、排课、数据请求)→ 检索依据(分发日志、手册、政策)→ 召回同校历史工单 → 白名单低风险动作自动执行(如推送音频版内容包)并记录 → 高风险动作(涉及身份、删除数据)转人工 → 草拟回复,置信 ≥ 90% 且在白名单内自动发送,否则待人审
成功标准
中位首次响应 ≤ 5 分钟;自动解决率有记录;数据请求 48 小时内闭环

三、界面语言与辅助语言

界面语言包覆盖中文、英文、斯瓦希里语、法语、西班牙语、荷兰语、印尼语七种,由运营端统一管理(设置、审核、修正、发布);未经审核的词条不进入生产界面。辅助语言(讲解脚手架)在 MVP 为英文,阶段 2 增加法语与斯瓦希里语;三国试点分别默认英文、斯瓦希里语、法语界面。

08 · Workstream D

落地实施Timeline, WBS, team, risk

本章给出以月份为单位的总体时间线、工作分解结构、团队编制、责任分配、预算结构、风险登记册、质量验收与试点运营手册。所有窗口以启动月为基准相对排列;若立项时间后移,全部窗口同步顺延。

总体时间线(启动月起 7 个月)

工作项9 月10 月11 月12 月1 月2 月3 月
A · 立项、候选校筛选、意向书与老师协议(4 周)
A · 试点合作备忘录与数据附件(3 周)
A · 教育部门、企业、渠道洽谈(滚动)
B · 三国需求调研与实地测量(每国 2–3 周)
C · 阶段 0 PoC(8 张测试卡,4–5 周)
D · MVP 开发与冻结(6 周)
D · 试点准备(每校 2 周)
D · 三国试点开课(每校 6 周,错开 ≤ 2 周)
D · 复盘与扩大决策(2 周)
D · 阶段 2:口语评测、对话、证书、企业班(约 12 周)
D · 原生安卓 App(10–12 周)
教研:HSK 一级课程与题库(持续)
合规:三国登记册、数据附件、同意书(前置后持续)
里程碑
里程碑时间窗口门禁(不满足不放行)
M1 PoC 门禁10 月上旬(启动后第 5 周)8 张测试卡有结论;各国直播 / 录播重心决定;ADR-1 至 7 落笔;三国合规备忘无「不可做」
M2 MVP 冻结10 月底(启动后第 8 周)功能范围锁定;老师 3 人完成培训;三校设备与同意书就位;HSK 一级前 6 周内容审校完成
M3 试点复盘12 月下旬(末校结课后 2 周内)§00 六条成功标准逐条有数据;单位成本模型有实测;三国对照结论;扩大试点名单确认
M4 阶段 2 验收次年 3 月口语评测与对话智能体上线;可验真结业证明发出;第一个企业委托班或付费席位落地

工作分解结构(WBS)

编号工作包主要任务负责窗口 / 时长
1.0项目启动立项决议、联合项目办公室、工具与文档库、事实卡制度协会负责人 + 平台负责人9 月,2 周
2.0对外洽谈候选校筛选、意向书、试点 MOU 与数据附件、老师协议、同意书本地化Ray · 刘老师 · 法务9–10 月,首轮 4 周
3.0需求调研三国访谈、问卷、实地测量、内容审计、PRD、画像、旅程图协会 + 平台产品;当地助教执行9–10 月,每国 2–3 周
4.0技术验证8 张 PoC、ADR、成本模型、MVP 范围平台技术9–10 月,4–5 周
5.0教研内容HSK 一级课程大纲、6 周课件与录播、题库 300 题、复习卡 400 张、审校流程刘老师 + 中国老师9 月起,MVP 前 6 周;持续
6.0MVP 开发租户与班级、录播管线、PWA 离线、伴学与复习智能体、拍照作业、老师端、学校端、运营端平台9 月中–10 月底,6 周
7.0试点准备学校设备与网络、缓存点、同意书回收、账号发放、老师与助教培训、支持渠道Ray + 试点校;老师培训由刘老师负责每校开课前 2 周
8.0试点运营三校各 6 周开课、周报、问题跟踪、数据采集、访谈回访联合项目办公室11–12 月
9.0复盘与扩大复盘报告、单位经济模型、三国对照、扩大试点名单、落地手册 v1协会负责人 + 平台负责人12 月,2 周
10.0阶段 2 建设口语评测、对话智能体、证书、原生 App、辅助语言、企业委托班平台 + 协会次年 1–3 月,约 12 周
11.0合规与安全三国登记册、数据附件、同意书、内容审核、审计留痕、渗透测试法务 + 平台安全贯穿

团队编制

角色人数来源投入职责
项目发起人1协会负责人每周 3 小时决策、资源、对外关系与签约
平台负责人(技术)1平台每周 1 天技术路径、架构决策、平台团队调度
项目经理1平台(或协会指定)全职计划、进度、风险、周会、文档、合同台账
教研负责人1刘老师半职课程体系、资源包、题库终审、老师招募与协调
非洲商务负责人1Ray半职三国候选校与当地伙伴、教育部门、支付渠道、市场伙伴
资源顾问2张老师、李社长按需教研评审与专家导学;出版与媒体
中文老师(中国侧)2–3协会聘每周 6–10 课时 + 备课录课、直播、复核、签发
运营 / 教务1协会半职排课、学员支持、家校沟通、数据登记
后端工程师2平台全职至 MVP 冻结,之后 1 人租户、课程、录播、同步、智能体接线
前端 / 移动工程师1–2平台全职PWA、老师端、学校端;试点后原生 App
AI / 智能体工程师1平台全职记忆、编排、评估集、Model Gateway 路由
设计1平台半职品牌落地、学员端界面
测试 / 运维1平台(兼)三分之一弱网测试、低端机测试、发布、监控
法务顾问1外聘(三国当地律师按需)阶段 0 约 8 天,之后按需三国法规备忘、数据附件、同意书、合同本地化
当地助教每校 1试点校每周 4 小时设备、点名、现场支持、测量

责任分配(RACI)

事项协会负责人平台负责人项目经理刘老师(教研)Ray(非洲商务)平台工程试点校法务
立项与预算ACRCCIIC
候选校选择与意向书AICCRICC
试点 MOU 与数据附件ACRIRCCR
PoC 结论与 MVP 范围CARCCRIC
课程、资源包与师资ICIA/RICCI
数据与合规设计ARCICRCR
试点运营与周报ICA/RRRCRI
教育部门与市场伙伴AICIRICC
复盘与扩大决策ARRCRCCI

A 批准 · R 执行 · C 咨询 · I 知会

预算结构(成本项与估算方法,不预设金额)

成本项驱动因子估算方法控本杠杆
协会人力老师课时、教研、运营课时费 × 课时 + 半职人员月薪录播复用,一位老师带三校
平台人力人月数(上表)按人月记录为项目投入复用既有能力,MVP 严格限范围
会议工具主持人席位数2–3 个付费席位月费先用现成工具,不买 RTC 分钟
CDN 与存储课程包大小 × 学员数 × 下载次数 × 国家数PoC-2 三国实测单价后建模低码率、学校缓存点、夜间下载
模型调用每学员每周会话数 × 每会话 tokenPoC-5 实测后建模,按租户设预算Gateway 按成本选路,缓存高频答疑
法务国家数 × 文书数按天计;三国各一位当地律师审校模板化同意书与数据附件
试点校支持设备补贴、流量补贴、助教津贴、缓存点设备按校按月共享设备、离线内容包
差旅候选校走访、开课到校按次;优先视频访谈Ray 一次行程覆盖多校

复盘时输出「每学员每月成本」三国实测值,作为此后与教育部门、企业与基金会谈价格的唯一依据。

风险登记册

#风险可能性影响应对负责人
R1跨境直播质量不达标(某国或全部)架构按录播为主设计;PoC-1 逐国决定重心;直播只保音频平台负责人
R2教研内容产能不足,MVP 无课可上9 月即启动,先做 HSK 一级前 6 周;复用现有课件;AI 辅助生成 + 老师审校刘老师
R3MVP 时间过紧,质量下降范围按本方案冻结;原生 App 后置;每周演示可用增量项目经理
R4某国候选校无法按期签约每国保留 1 所备选;开课错开窗口 ≤ 2 周;最坏两国先行Ray
R5试点校设备或供电不足调研先摸底;共享设备与多账号;离线内容包;缓存点Ray
R6数据出境或未成年人合规未闭环PoC-8 前置;三语同意书;数据最小化;不合规不开课法务
R7智能体出错未被发现,损害教学信任有出处 + 置信不足即转人工;每周抽样 50 条人工评审;错误公开记录AI 工程
R8老师抵触或流失老师参与设计;工具减负而非监控;课时与音色授权有报酬刘老师
R9模型与流量成本失控租户预算与降级;月度成本报表AI 工程
R10「汉桥」名称与「汉语桥」近似引发争议汉桥只作内部代号;对外统一教育品牌另行命名并做商标检索协会负责人
R11对外材料数字无法核证事实卡制度;未核证数字不上线协会内容负责人
R12单点依赖某一云或模型供应商云中立矩阵与 Model Gateway;每个第三方有备选平台负责人
R13三国法务本地化拖延签约先签非约束性意向书;MOU 采用统一英文主文 + 各国附件法务

质量与验收

  • 每周演示:周五 30 分钟,用真实学员账号在弱网模拟(限速 256 kbps、丢包 5%)下演示本周增量。
  • 低端机基线:团队常备 2 台 2 GB 内存安卓机,所有学员端功能在其上验收。
  • 智能体评测集:课程内 200 题评测集,每次发布跑一遍,有出处率、错误率、转人工率三项不退步。
  • 安全:MVP 冻结前做一次渗透测试与依赖扫描;租户隔离有自动化测试。
  • 用户验收:老师、学员、校方各出一份签字的验收单,内容为「能否独立完成一周的教与学」。

试点运营手册(每校)

开课前两周

  • 学员名单与分组(按年龄与基础);同意书回收率 100%
  • 账号与 PIN 发放,设备安装与离线包预下载;缓存点部署
  • 老师与助教各 2 小时培训,含降级流程演练
  • 支持渠道:WhatsApp 群 + 站内助手 + 工单;值班表

每周节奏

  • 周一录播与练习包发布
  • 周三直播(或音频直播)30 分钟,缺席看回放
  • 每天 15 分钟智能体陪练与复习卡
  • 周五老师复核队列清零,周报(出勤、作业、准确率、成本、故障)
  • 每两周一次学员与家长回访

关键指标(试点仪表盘)

维度指标采集方式
参与周活跃学员比例、出勤率(含回放)、每周练习时长中位数系统日志
学习掌握度周增量、模拟测验通过率、作业提交率知识图谱与评测记录
老师杠杆每班每周老师投入课时、复核队列处理时长、转人工比例、每课时服务学员数老师端日志 + 工作日志
可信智能体抽样准确率、有出处率、错误被发现前的暴露次数每周人工评审
可达离线练习占比、同步成功率、录播完播率、降级触发次数(按国)客户端遥测
成本每学员每月:模型、流量、会议工具、人力分摊(按国)月度成本表
合规同意书覆盖率、数据请求响应时长、安全事件数登记册
09 · MOU & contracts

备忘录与合同Instruments, key clauses, reference samples

本章列出落地所需的全部对外法律文书、每份文书的关键条款与参考样本。样本以英文主文为准、中文为对照,供协会法务与各国当地律师本地化时参考;本章不构成法律意见。所有签约主体在正式签署前标注「待定」。

一、文书清单与签署顺序

#文书签署方(待定)性质时点语言
C1试点校意向书(Letter of Intent)协会 ↔ 试点校非约束性(保密条款除外)9 月,候选校确定后英 / 法
C2试点合作备忘录(Pilot MOU)协会 ↔ 试点校(及其所属教育机构,若适用)约束性条款:保密、数据保护、知识产权、终止;其余为意向10 月,开课前英文主文 + 当地语言译本
C3数据保护附件(Data Protection Schedule)随 C2 签署;平台作为处理者出具处理者承诺约束性随 C2英文
C4监护人 / 学员同意书(Consent Form)监护人(未成年)或学员本人(成年)→ 协会约束性开课前,覆盖率 100%英 / 法 / 斯瓦希里语
C5远程中文老师服务协议(Teacher Services Agreement)协会 ↔ 老师个人约束性9 月起中文
C6音色复刻授权书(Voice Licence)老师 → 协会(平台为技术执行方)约束性,单独签署,可单独撤回随 C5 或阶段 2中文
C7当地助教协议(Local Facilitator Agreement)协会或试点校 ↔ 助教约束性开课前英 / 法
C8企业委托班服务协议(Corporate Programme Agreement)协会 ↔ 企业约束性阶段 2英 / 中
C9教育部门合作框架协议(Framework of Cooperation)协会 ↔ 教育主管部门意向性为主试点出数据后英 / 法

二、合同管理流程

  1. 模板:本章样本为统一模板;协会法务定稿英文主文。
  2. 本地化:每国一位当地律师审校数据保护、未成年人、劳务与税务条款,输出「本地化差异表」。
  3. 签约方确认:由 Ray 确认学校法人主体与授权签字人;由协会负责人确认协会签字人。
  4. 签署:电子签为主(DocuSign 或同等工具);当地要求纸质时以扫描件为准并留存原件。
  5. 台账:项目经理维护合同台账(文书、版本、签署日、期限、到期提醒、数据附件版本)。
  6. 变更:范围、期限、数据用途的变更以补充协议形式签署,并同步更新合规登记册。

三、关键条款与参考样本

C2 · 试点合作备忘录(结构与样本条款)

1 目的与范围
在 [学校] 开展为期 6 周、20–40 名学员的中文学习试点;内容为 HSK 一级前 6 周;通过 learn.gacee.org 平台交付。
2 期限
自签署日至试点复盘完成;任一方可提前 30 天书面通知终止;终止不影响第 7、8、10 条。
3 协会义务
提供课程内容、远程老师、平台访问、培训与支持;试点期免收任何费用;出具学习报告与结业证明。
4 学校义务
组织学员、提供固定课表与场地、指定助教、协助回收同意书、允许在校内进行网络测量。
5 费用
试点期免费;学校不承担系统费用;协会不向学员或家长收取费用。
6 知识产权
课程内容、教材、品牌归协会;学员作品归学员本人,协会在获同意后可用于教学与非商业展示;平台软件为协会之技术服务提供方所有,学校获得试点期使用许可。
7 数据保护
依附件 C3;协会为控制者,学校为本校学员数据的联合控制方;跨境传输仅在取得监护人同意并符合当地法规的前提下进行。
8 保密
双方对未公开的试点数据、内容与商务信息保密;期限至终止后 2 年。
9 品牌与宣传
任何一方对外提及本试点须事先书面同意;学生肖像与作品使用须单独同意。
10 责任与争议
本备忘录除第 6–9 条外不构成有约束力的义务;争议先协商,协商不成提交 [新加坡国际仲裁中心 / 当地法院,待法务确定]。
11 试点后安排
复盘后双方可另行签署正式合作协议;学校可要求导出或删除本校学员数据。
Sample clause · Purpose and non-binding nature

This Memorandum of Understanding records the intention of the Global Association of Cultural and Educational Exchange (“the Association”) and [School] (“the School”) to conduct a six-week Chinese-language learning pilot for 20 to 40 learners using the Association's learning platform. Except for Clauses 6 (Intellectual Property), 7 (Data Protection), 8 (Confidentiality) and 9 (Publicity), this Memorandum does not create legally binding obligations. No fee is payable by the School, the learners or their guardians during the pilot.

本备忘录记录协会与 [学校] 就六周中文学习试点的合作意向;除知识产权、数据保护、保密、宣传四条外不产生约束力;试点期学校、学员与监护人不承担任何费用。

C3 · 数据保护附件(要点)

条款内容
角色协会为数据控制者;学校为本校学员数据的联合控制方;技术平台为处理者,仅按协会书面指示处理
数据项学员:姓名、年龄段、学校与班级、学习记录、作业与录音;监护人:姓名与联系方式;助教:姓名与联系方式。不采集身份证件号、精确住址、生物特征
目的教学交付、学习评估、学习报告与证明、试点研究(脱敏汇总);不用于广告;试点期不用学员数据训练任何通用模型
存储与传输核心数据存新加坡;传输与静态加密;跨境传输依据当地法规的合法路径与监护人同意
保留试点结束后 12 个月或账号注销后 30 天(以先到者为准)删除;学校可要求提前导出与删除
子处理者清单列明(云、CDN、会议工具、模型服务);变更前 30 天通知
数据主体权利访问、更正、导出、删除、撤回同意;48 小时内响应,30 天内完成
安全事件发现后 72 小时内通知学校与监管机构(按当地要求)
审计学校可每年一次要求提供隔离测试与安全测试摘要

C4 · 监护人 / 学员同意书(要点,三语)

  • 收集的数据项与目的;数据存放在新加坡;跨境传输的说明与法律依据。
  • 直播录制与语音练习会被记录用于教学与评估;不用于广告或训练通用模型。
  • 随时撤回同意与要求删除的方式;联系人与响应时限。
  • 照片、录音与作品用于协会宣传或出版的单独勾选项,默认不勾选。
  • 成年学员本人签署;未成年学员由监护人签署,学校留存副本。
Sample clause · Consent to recording

I understand that live classes are recorded and that my child's voice recordings from pronunciation practice are stored securely in Singapore and used only for teaching and assessment. I may withdraw this consent at any time by contacting [email], after which the recordings will be deleted within 30 days. ☐ I additionally agree that my child's photographs or written work may be used in the Association's non-commercial publications (optional).

C5 · 远程中文老师服务协议(要点)

条款内容
服务内容录课、直播、复核队列、签发关键判定、参与教研;每周课时上限与备课时数
报酬课时费 + 录播一次性费用 + 复核按件;月结;税务按老师所在地法规
录播版权录播内容著作权归协会,老师享有署名权;协会可在平台内向学员播放,不对外销售;合作终止后录播可继续用于已开班级至结课
教学判断老师对批改、签发、证书具最终决定权;平台工具仅提供建议
个人信息按最小化原则处理老师个人信息;老师数据存储与出境依中国 PIPL;不公开老师联系方式
保密与行为规范学员数据保密;未成年人保护行为规范;禁止私下收费
终止任一方提前 30 天通知;在途班级完成交接

C6 · 音色复刻授权书(要点)

  • 授权范围限于:课程旁白、发音示范、图书馆朗读、纠错反馈;不用于其他用途。
  • 授权可随时撤回,撤回后 30 天内停止生成并删除音色模型;已生成的课程音频在已开班级结课后删除。
  • 授权有独立对价;音色模型仅存于新加坡,不对外提供。

C8 · 企业委托班服务协议(要点)

条款内容
服务班型(人数、周数、目标场景)、交付方式(录播 + 直播 + 智能体陪练)、HR 报表频率、结业证书
费用与结算按班报价(待定);开班前预付 50%,结业后付清;发票由协会开具
数据员工学员享有与学校学员同等数据主体权利;HR 仅见出勤、进度与结业状态;企业不得要求作答明细
服务水平试点期以尽力为准并记录可用性;正式期月可用性 99.5%,24 小时响应
知识产权课程内容归协会;企业定制内容(术语表、场景)归企业,协会获教学使用许可
终止提前 30 天通知;已开班按比例结算

C9 · 教育部门合作框架协议(要点)

  • 合作范围:课程与当地课程体系衔接、试点校扩展、认证路径探讨、教师培训。
  • 数据主权:本国学员数据依本国法规处理;教育部门可获脱敏汇总报告。
  • 费用:普惠版对公办学校免费;其余另行商定。
  • 性质:框架性、非排他;具体项目另签。

四、样本使用说明

法律声明

本章样本仅供协会法务与当地律师起草时参考,不构成法律意见;各国数据保护、未成年人、劳务与税务条款须经当地律师审校后方可签署。所有金额、比例与期限在正式文本中按谈判结果填写。

10 · Pedagogy & content

教研与内容Content is engineering too

课程内容不能由技术生成替代教研。教研是本项目最易成为瓶颈的人力投入,须与技术同期启动。三国试点使用同一套内容,仅辅助语言不同。

阶段内容目标产出审校
MVP(10 月底)HSK 一级前 6 周:拼音与声调、150 词、基础句型、日常对话 6 个场景6 周课件与录播、300 道客观题、复习卡 400 张、6 次拍照作业题、英文脚手架中国老师初审 → 刘老师终审 → 入库
试点期(11–12 月)HSK 一级全部 + 一级模拟考12 周内容、模拟考 2 套;法语与斯瓦希里语脚手架初稿同上,加试点老师与助教反馈修订
阶段 2(次年 1–3 月)HSK 二级、口语场景 12 个、法语 / 斯瓦希里语讲解脚手架定稿课程知识图谱扩到二级、口语题库增加母语审校人
  • 标准对齐:以 HSK 3.0(三等九级)与 YCT 为锚,每个知识点在图谱中标注等级与来源;平台结业证明与官方认证严格分开。
  • 内容生产方式:老师撰写大纲与关键讲解,AI 辅助生成练习与变式,老师逐题审校后入库;未经审校的内容不进入图谱。
  • 与出版和未来书院衔接:课程模块按未来书院四大板块编码,优质讲义与学生作品沉淀为出版素材;作文大赛题目作为写作练习来源;出版新书自动上架图书馆。
  • 与当地体系衔接:试点期收集三国课程标准与学期节奏,为阶段 3 与教育部门谈认证做准备。
11 · Compliance

合规与数据Built into the foundation

合规登记册逐国、逐义务、逐证据,由法务确认后才作为设计输入。下表列出需确认的边界,不构成法律意见;三国的具体法规适用性以 PoC-8 法务备忘为准。

范围法规(需法务确认适用性)主要义务系统落点状态
新加坡(协会与核心数据)PDPA同意、目的限制、保护、保留、数据出境、数据保护官隐私政策、同意管理、新加坡驻留、导出与删除接口平台已对齐;协会需指定数据保护官
尼日利亚(学员)NDPA 2023 与 NDPC 实施规则合法基础、未成年人由监护人同意、跨境传输条件、可能的登记义务监护人同意书、数据最小化、出境记录、学校子租户隔离待法务备忘
肯尼亚(学员)Data Protection Act 2019(ODPC)控制者 / 处理者登记、未成年人保护、跨境传输须 ODPC 认可的保障措施同上;登记状态写入登记册待法务备忘
科特迪瓦(学员)2013 年个人数据保护法(ARTCI)处理申报或授权、未成年人保护、跨境传输授权同上;申报文件归档待法务备忘
中国(远程老师)PIPL 与数据出境相关规定个人信息最小化、告知同意、出境评估门槛老师数据最小化;录播与音色授权书;不在中国侧落地媒体节点待法务备忘
欧洲华校(阶段 3)GDPR数据处理协议、数据主体权利DPA 模板、导出与删除阶段 3
内容与未成年人保护各国儿童在线保护要求内容审核、举报与申诉、直播录制告知审核队列、举报入口、录制提示(每课开始三语播报)MVP 内置
AI 治理新加坡 AI 治理框架(自愿)透明、人类监督、可追溯智能体输出有出处、转人工、审计留痕、错误公开记录设计原则已覆盖

数据流与主权(写入数据保护附件)

  • 协会为数据控制者,技术平台为处理者,学校为各自学员数据的联合控制方;三方在备忘录与附件中写明。
  • 核心数据存新加坡;学员端本地数据加密;情节记忆只追加、可导出、可删除;账号注销后 30 天内删除。
  • 学校之间数据真正隔离:租户级隔离测试是发布门禁的一部分。
  • 试点期不做任何数据商业化,不用学员数据训练通用模型。
  • 三国各设一名当地数据联系人(可由助教兼任),负责数据主体请求的首次接收。
12 · Governance

治理与沟通Joint programme office

联合项目办公室

协会负责人、平台负责人、项目经理、教研负责人(刘老师)、非洲商务负责人(Ray)五人组成;每周一 45 分钟例会;项目经理维护唯一的计划、风险表与合同台账。

决策权限

范围与时间变更:项目办公室;预算与对外协议:协会理事会;架构与技术选型:平台负责人;课程内容与师资:刘老师;三国试点校与当地关系:Ray。

文档体系

一个共享云盘:计划、PRD、ADR、合规登记册、合同台账、周报、会议纪要、验收单。所有对外材料从品牌手册与事实卡取。

沟通频率参与形式
项目例会每周一项目办公室进度、风险、决策,45 分钟
周五演示每周五全体 + 试点校代表(可选)真实账号弱网演示,30 分钟
试点周报每周五理事会、三校、老师一页:指标、问题、下周计划;按校分发
理事会汇报M1 – M4 各一次理事会门禁结论与决策请求
试点校沟通每两周校长、助教、家长代表回访电话或到校
变更控制随时项目经理受理变更单:影响时间 / 成本 / 合规者须项目办公室批准
13 · Decisions

决策清单What must be decided, by whom, by when

#决策建议决策人时限
D1立项与对外签约授权理事会决议授权协会负责人签署 C1–C7协会理事会9 月第 1 周
D2三个试点国家尼日利亚、肯尼亚、科特迪瓦;备选加纳、坦桑尼亚、塞内加尔协会理事会9 月第 2 周
D3每国一所试点校按四维打分短名单,Ray 推荐协会负责人9 月第 4 周
D4MVP 范围(含原生 App 后置)按 §05 表项目办公室M1
D5各国直播 / 录播重心按 PoC-1 逐国结论平台负责人M1
D6中国侧是否落地媒体或接入节点试点期不落地平台负责人 + 法务M1
D7统一教育品牌名「同文」或「书桥」,商标检索后定;「汉桥」仅作代号协会理事会9 月底
D8试点期收费全部免费;阶段 2 起按席位与委托班收费协会负责人M1
D9争议解决地与适用法新加坡法 + 新加坡国际仲裁中心,当地律师确认可执行性协会负责人 + 法务10 月,C2 签署前
D10扩大试点校名单与国家按三国对照结论与四维打分协会理事会M3
14 · Appendix

附录 · 模板Ready to use

A. 校长访谈提纲(60 分钟)

  1. 学校概况:年级、人数、中文课现状、每周课时与时段。
  2. 设备与网络:学校可用设备数量与状况、网络运营商与资费、停电频率、是否有可共享的教室与投影、能否放置缓存点设备。
  3. 老师与助教:谁在课堂上组织学生,母语与英语 / 法语 / 斯瓦希里语使用比例。
  4. 家长:沟通渠道、付费习惯与能力、对孩子数据的态度。
  5. 成果:家长和学校最看重什么(证书、考试、能开口);对 HSK / YCT 的认识。
  6. 顾虑:对 AI 的担心、对数据的担心、对时间与成本的担心。
  7. 合作:学校愿意提供什么、希望协会提供什么、免费试点期的期待、试点后的意愿。

B. PoC 记录表字段

测试卡编号 · 国家与学校 · 日期与时段 · 地点与运营商 · 设备 · 执行人 · 步骤 · 原始测量值 · 是否通过 · 证据文件(截图 / 日志 / 录音)· 结论与建议 · 复核人。

C. 试点周报模板(一页,按校)

本周数字(活跃、出勤含回放、练习、作业、准确率、转人工、成本)· 三件做得好的事 · 三个问题与负责人 · 下周计划 · 需要理事会或校方的决定。

D. 候选校打分表

国家 · 学校 · 联系人 · 网络与供电(1–5)· 需求与组织力(1–5)· 合作方成熟度(1–5)· 合规可行(1–5)· 总分 · 走访日期 · 结论(签约 / 备选 / 排除)。

E. 合同台账字段

文书编号 · 文书类型 · 签署方 · 国家 · 版本 · 签署日 · 生效日 · 到期日 · 数据附件版本 · 本地化差异表 · 保管位置 · 到期提醒。

F. 术语

PoC
概念验证,用最小实验回答一个风险问题
MVP
最小可用产品,能让一个真实班级完整用起来的最小功能集
PWA
可安装的网页应用,可离线运行
RTC
实时音视频
TIDAR KAG
平台自研的有出处知识增强生成,用于低幻觉答疑与掌握度判定
Agent Harness
智能体的构建、运行、评估与治理底座
MOU / LOI
合作备忘录 / 意向书
DPA
数据处理协议或数据保护附件
ADR
架构决策记录
HSK / YCT / HSKK
汉语水平考试 / 中小学生汉语考试 / 汉语口语考试
GACEE · 汉桥落地方案 v3.0 · 2026 年 9 月所有指标阈值为工作假设,实测后修订;签约方待定
GACEE · Global Association of Cultural and Educational Exchange

HanQiao
Delivery Plan汉桥项目落地方案 · Three-country pilot · v3.0

This plan sets out how HanQiao, the Association's AI-enabled Chinese-learning platform, is to be delivered in three countries with one school in each, and how the pilot data will serve as the basis for the next round of external partnerships. It follows four parallel workstreams — external engagement, needs research, technical validation and delivery — and includes application scenarios, the technical route, and reference samples for memoranda and contracts. Time windows are expressed in months and required durations only; every signatory is marked “to be confirmed” until execution; every unverified assumption is flagged in the text.

Version 3.0 · September 2026Basis: GACEE Association profile · JE HanQiao architectureClassification: Association internal · external edition to follow
00 · Executive summary

Executive summary执行摘要

The core task is to place remote Chinese teachers, orchestrable agents and an offline-first learner client into the regular timetables of three schools, complete six weeks of teaching under weak-network, power-cut and shared-device conditions, and obtain verifiable operating data.

The project in one sentence

Teachers first; three countries and three schools as the proving ground; weak network and offline by default; compliance in the foundation. The public entry point is learn.gacee.org and certificate verification is verify.gacee.org.

Objectives

Approve the project, shortlist schools and start technical validation in September; freeze the MVP by the end of October; open six-week classes in one school per country from November; complete the review and decide on expansion before year end.

Pilot footprint

Three countries × one school each × 20–40 learners per school. The three countries cover English, Kiswahili and French as scaffolding languages and three data-protection regimes, validating portability on one system.

Delivery roles

The Association (content, teachers, international resources, programme design) + pilot schools (learner organisation and on-site delivery) + remote Chinese teachers (teaching and sign-off) + the technology platform (systems and AI). Signatories are confirmed after legal review in each country.

Eight principles

Teacher-led; no hallucination; inclusion first; weak-network and offline by default; mother-tongue scaffolding; compliance in the foundation; light before heavy; aligned with standards and local systems.

Four parallel workstreams and their windows

WorkstreamWindowDurationDeliverablesOwner
A External engagementFrom September, ongoingFirst round 4 weeks, then rollingShortlist of candidate schools in three countries; pilot MOU and letters of intent (signatories TBC); teacher agreements; guardian consent templatesAssociation leads; Ray for Africa; Liu Laoshi for teachers
B Needs researchSeptember–October2–3 weeks per country, in parallelRequirements document v1; five personas; field network and device survey for three countries; compliance register v1Association and platform jointly; local partners run measurements
C Technical validation (Phase 0)September–October4–5 weeksAll eight PoC cards concluded; live-first or recording-first decided; ADR-1 to 7 writtenPlatform
D Delivery (Phases 1–3)From mid-SeptemberMVP build 6 weeks; pilot preparation 2 weeks; 6 weeks per school; review 2 weeks; Phase 2 about 12 weeksMVP freeze → three schools open → review report and unit cost → Phase 2 acceptanceJoint programme office
Prior judgement

Bringing a real class into use within two months is feasible only if the MVP is defined as recorded lessons + live classes on existing meeting tools + offline practice in an installable web app (PWA) + two agents (study companion and review). A native offline Android app needs 10–12 weeks to mature; it starts after the pilot opens and ships as a beta in Phase 2. This is not a reduction in scope: the weak-network-and-offline principle is delivered on the web first and migrated to native afterwards.

Success criteria (accepted item by item at the pilot review)

  1. One school per country, 20–40 learners each, completes six weeks; attendance (including replays) and homework submission are recorded by the system.
  2. Teachers' weekly hours fall while learner practice rises, both evidenced by logs; the number of learners served per teacher hour is measured.
  3. Sampled accuracy of agent answers and the hand-off rate are measured; no incorrect judgement goes undetected.
  4. Learners complete practice offline during outages and sync afterwards with zero data loss; network measurements are comparable across the three countries.
  5. School data isolation, guardian consent and cross-border transfer are each evidenced in writing in all three countries.
  6. A measured cost per learner per month is produced and becomes the sole basis for pricing discussions with education authorities, employers and foundations.
01 · Landscape & roles

Landscape and roles落地格局与角色

The long-term four-part structure

Publishing

Curates quality content: competition anthologies, course textbooks, cultural readers; in turn supplies the platform library.

Future Academy

Builds courses and real educational settings across four strands: Chinese culture, world civilisations, AI and the future, growth in practice.

AI platform (HanQiao)

Scales content and teaching capacity: remote teachers + agent orchestration + offline learning + online library.

The Association

Connects schools, teachers, experts, international resources and programmes; the hub and brand holder of the four.

HanQiao offers three public entry points under the Association's domain: the platform at learn.gacee.org, certificate verification at verify.gacee.org, and single sign-on with the Association website from Phase 2.

Delivery roles and responsibilities

RoleResponsible forOwns / decidesDoes not
The Association (GACEE)Educational content, teacher resources, international resources, programme design, relationships with schools and education authorities, brand, external contractingBrand, course copyright, learner relationships; as data controller, decides data purposesBuild its own engineering team
Pilot school (one per country, TBC)Learner organisation, timetable slots, venue and devices, local facilitator, parent communication and consent collectionManagement of its own learners; joint control of its own learner dataBear system costs; the pilot is free
Remote Chinese teachers (China side)Recording, live classes, review queue, sign-off of key judgements, participation in curriculum workFinal say on pedagogical judgements; scope of voice and recording licencesOperate systems; handle sensitive learner data
Local education authority (engaged from Phase 2)Access, curriculum alignment, certification path, public-school channelNational accreditation and accessIntervene in daily operations during the pilot
Technology platform (service provider)Systems, AI capability, operations, security and compliance engineering; processes data on the Association's instructionsTechnical implementation of the platform and agent baseAct in the Association's name; charge schools directly
Corporate clients (from Phase 2)Commission classes for local staff; provide rosters and an HR contactUse of reports and certificates for its own learnersAccess other tenants' data

Core roles and resource people

RolePersonResponsible forPosition in this plan
Programme sponsorAssociation leadDecisions, budget, external relations, coordination of the four-part structureApprover at each milestone; authoriser of external contracts
Africa business resources leadRayCandidate schools in three countries, local partners, education authorities, payment and local channels, African market partnersWorkstream A (Africa); school-side interface for pilot preparation and operations
Curriculum, resource packs and teachers lead (pedagogy lead)Liu LaoshiChinese curriculum, content packs, recruitment, training and scheduling of China-side teachers, final content review§10 pedagogy, teacher agreements, MVP content gate
Education expert resourcesZhang LaoshiCurriculum review, expert guidance, principal and school networks (division of work to be confirmed by the Association)Curriculum review and principal-network support
Publishing and media resourcesPublisher LiAnthology and textbook publishing, media coverage, content editing (division of work to be confirmed)Library sources; the “publishing” strand
Platform lead (technology)PlatformTechnical route, architecture decisions, platform team allocationTechnical approver for Workstreams C and D

External revenue and funding sources (consistent with the Association's non-profit status)

  • Institutional seats: schools, Chinese-community schools and training centres purchase per-learner seats with a free basic tier and per-seat upgrade to the full tier. Pricing is to be determined and is not assumed here; all three pilot schools are free.
  • Corporate commissioned classes: Chinese enterprises abroad commission classes for local staff, settled per class, with completion certificates and learning reports (from Phase 2).
  • The Association's own income: event registration, study-journey fees, competition entry fees, membership fees, collected through the Association's own accounts.
  • Philanthropic funding: EduLink-Africa and public lectures supported by foundations and donors; the inclusive tier is free for ordinary schools. The review report is the core material for funding applications.
02 · Three-country pilot

Three-country pilot design三国试点设计

Three countries are chosen to validate, with the smallest possible sample, that one system is portable across different scaffolding languages, network profiles and data-protection regimes. One school per country preserves depth of engagement and data quality in each class.

Candidate countries (recommended, subject to the Association's confirmation; schools and signatories TBC)

SlotRecommended countryRationaleScaffolding languageData law (legal confirmation required)Alternate
P1NigeriaExisting online Chinese classes and school relationships; the largest anglophone West African market; power and mobile-network conditions test weak-network and offline design most severelyEnglishNDPA 2023 and NDPC implementation rulesGhana
P2KenyaEast African hub and natural test bed for Kiswahili scaffolding; submarine cable and 4G coverage better than West Africa, serving as the “better network” control groupKiswahili + EnglishData Protection Act 2019 (ODPC)Tanzania
P3Côte d'IvoireEconomic centre of francophone West Africa; validates the French interface and French scaffolding and provides the basis for Phase 3 expansion in francophone countriesFrench2013 personal data protection law (ARTCI)Senegal · Cameroon

Should the Association prefer one slot outside Africa to test an Asian market, Indonesia is an alternate (the Indonesian interface pack is ready); keeping all three in Africa simplifies Ray's coordination and makes the review comparable.

Four-dimension school scoring (1–5 per dimension; total ≥ 14 enters the shortlist)

Network and power

Hours of usable school Wi-Fi or 4G, frequency of power cuts, feasibility of a local cache point.

Demand and organisation

An existing Chinese class or clear demand, a fixed six-week timetable, a teacher willing to act as facilitator.

Partner maturity

Short decision chain at principal level, open parent channels, willingness to sign an LOI and run the consent process.

Compliance feasibility

Guardian consent workable, a lawful route for cross-border transfer, no “cannot do” conclusion.

Uniform specification per school

ItemSpecificationNote
Learners20–40, in 1–2 groups by age and levelFewer than 20 is an insufficient sample; more than 40 overloads the teacher review queue
Duration6 weeks; per week: one 30-minute live class + one recorded lesson + 15 minutes of practice dailyStart dates across schools staggered by no more than 2 weeks so one teacher can be scheduled
TeachersOne China-side lead teacher per school; one preparation-and-review teacher shared across the threeRecordings reused; one teacher serves several schools
DevicesLearners' own phones (including shared) + 2–4 shared school devices + one cache-point device (optional)A 2 GB-RAM Android phone is the acceptance baseline
Local facilitatorOne per school, 4 hours per weekRoll call, devices, on-site support, network measurement
FeesFree throughout the pilot; the Association bears third-party direct costsDevice and data subsidies recorded per school per month into the unit-cost model
OutputsLearning report, completion certificate (issued by the Association, verifiable), school weekly report, term report for the education authorityCertificates are strictly separate from official HSK results

Cross-country comparison design

All three schools use the same content (the first six weeks of HSK Level 1) and the same system; only three variables differ: scaffolding language, network profile and legal regime. The review compares results across these variables to establish which capabilities are portable and which require per-country adaptation — the basis for Phase 3 scaling.

03 · Workstream A

External engagement对外洽谈

Engagement proceeds “internal before external, near before far”: the plan, pilot specification and partner requirements are finalised first, and only a mature proposal is taken outside. All external commitments derive from the fact sheet; every signatory is marked “to be confirmed” until legal review.

Stakeholder map

CounterpartyWhat they wantWhat we needWhenOwner
Association boardPublic-benefit fit, controlled risk, brand valueApproval, budget frame, authority to contractSeptember, week 1Association lead
Candidate schools (2–3 per country; one signed per country)A more stable Chinese class, visible learner outcomes, no added cost, data securityLearners, venue, network and device sharing, facilitator, consent processFrom September; LOI within 4 weeksRay + Association lead
Remote Chinese teachers (China side)Reasonable hours and pay, usable tools, no replacement by AI, clear recording and voice licencesParticipation in curriculum work and the pilot; sign-off of key judgementsFrom SeptemberLiu Laoshi
Local facilitators / partnersStipend, training, recognitionOn-site execution, measurement, roll callAfter LOIRay
Education expert (Zhang Laoshi) · Publishing and media (Publisher Li)Professional standard and influence; publishable outcomes and coverage materialCurriculum review, expert guidance; anthology and textbook publishing, media coverageFrom SeptemberAssociation lead
Local education authoritiesFit with local curriculum and certification, data complianceAccess, certification alignment, public-school channelOnce pilot data exists (from December)Ray + Association lead
Chinese-community schools / international schools / university Chinese departmentsQuality courses and test-centre cooperationChannels and test centresPhase 2Association
Chinese enterprises abroadStaff Chinese training with reports and certificatesCorporate commissioned classesAfter MVP stabilises (from December)Association
Foundations / donorsMeasurable public-benefit outcomesFunding for the inclusive tierAfter the review reportAssociation

Sequence and the objective of each step

September · week 1 · internal alignment
The board approves the project, budget frame, brand and domain resolutions, and authority to contract. Output: board resolution.
September · weeks 1–4 · school shortlist and LOIs
Ray produces a shortlist using the four-dimension score, visits or interviews 2–3 schools per country, and signs one LOI per country; consent-form localisation starts in parallel. Output: 3 LOIs; trilingual consent template.
September · from week 2 · teacher agreements
Liu Laoshi signs hourly agreements with the first 2–3 China-side teachers, covering recording copyright, voice-cloning licence (separate instrument) and confidentiality. Output: 3 teacher agreements; voice licences.
October · pilot MOU
Pilot MOU + data-protection schedule signed with each of the three schools (and their governing body where applicable); samples in §09. Output: 3 MOUs; 3 data schedules.
From December · education authorities and channels
Pilot data (attendance, homework, accuracy, cost, three-country comparison) taken to education authorities and community schools. Output: draft framework of cooperation.
From December · corporate classes
Proposals to 2–3 Chinese enterprises; one paid pilot class first. Output: corporate programme agreement (sample in §09).

Engagement materials

  • Association profile: condensed to 15 pages in the brand-manual visual system.
  • HanQiao one-pager: positioning, eight principles, three-country pilot specification, what partners provide and receive.
  • School edition (English / French / Kiswahili): what learners do each week, what the school provides, how data is protected, how long the free period lasts, what follows.
  • Education-authority edition: alignment with HSK standards, path to local curriculum integration, data-sovereignty commitments, sample term report.
  • Corporate edition: commissioned-class process, HR report sample, certificate sample, per-class pricing (TBD).
  • Demonstration environment: role demo accounts on learn.gacee.org (learner, teacher, school, enterprise, operations); replaced by an anonymised view of a real class once the pilot opens.
  • Fact sheet: the single source for every external figure, with provenance and update date.
Engagement discipline

No verbal commitments on metrics (latency, accuracy, country coverage); all external figures come from the fact sheet; questions involving minors or data are answered “designed to local law; written reply after legal confirmation”; no party commits on behalf of another.

04 · Workstream B

Needs research需求调研

The research establishes the actual learning conditions of teachers and learners in the three pilot schools: available devices, weekly time slots, modes of study and network environment. Its conclusions determine the MVP scope. The three countries run in parallel, 2–3 weeks each, with one toolkit, so results are directly comparable.

Subjects and samples (per country)

SubjectSampleMethodQuestions to answer
Principal and academic office2Structured 60-minute interview (Appendix A)Timetable slots, venue, device count, power supply, parent channels, payment habits, certification needs
Local class teachers / facilitators2Interview + one lesson observationClassroom organisation, level distribution, use of mother tongue and English / French / Kiswahili, concerns about AI
Remote Chinese teachers3 (shared across countries)Interview + one-week work logPreparation and marking time, cross-border network experience, tasks they most want relieved of
Learners15, by age bandQuestionnaire (local language) + focus group of 6Phone ownership and sharing, daily study time and slots, data budget, motivation
Parents10Questionnaire + phone interviewWillingness and ability to pay, attitude to children's data, expected outcomes (certificate / score / speaking)
Existing Chinese coursesThe Association's courses and the school's materialsContent auditExisting courseware, progress, assessment, reusable content assets

Field measurement toolkit (run by local facilitators, one template for all three countries)

Network measurement

  • At three times of day (morning, afternoon, evening) measure downlink / uplink / latency / packet loss; record operator and technology (2G / 3G / 4G / 5G).
  • One 20-minute Zoom audio call to Singapore; record drops and subjective clarity (1–5).
  • Download a 50 MB test file; record time and interruptions; test Cloudflare, Alibaba Cloud CDN and direct object storage separately.
  • Record power-cut hours over one week and any backup power.

Device and payment survey

  • Learner and school devices: brand and model, Android version, RAM, free storage, sharing.
  • Shared tablets / computers / projector / speakers; whether an always-on device can act as a local cache point.
  • Monthly data plan cost and usage; common payment methods (mobile wallet / cash / bank transfer).
  • Local and teaching languages; learners' reading level in English / French / Kiswahili.

Deliverables

DeliverableContentDue
5 personasRemote teacher, local facilitator, learner (child / adolescent / adult), parent, school administrator; differences noted per countryEnd of research week 2
3 journey mapsA learner's week (online / weak network / offline), a teacher's week, a school from enrolment to completionEnd of research week 2
Three-country field reportMeasured network, device, power and payment data with raw records; cross-country comparison tableEnd of research week 3
Requirements document v1Functional (MoSCoW), non-functional (weak network, offline, low-end devices, multi-account, scaffolding languages), data and compliance requirementsEnd of research week 3
Compliance register v1Law, obligation, owner and evidence per country (§11)End of research week 3

Research and validation run in parallel: first-week network measurements feed PoC-1 and PoC-2 directly.

05 · Workstream C

Technical validation and Phase 0技术验证与阶段 0

Architecture, the Agent Harness, JE Engine orchestration and TIDAR KAG are described in §06; this chapter defines validation methods and pass criteria. MVP freeze requires all eight cards concluded. Network-related cards run once per country; the others run once.

Reuse and build (summary)

Reused directly (in production, verified)

  • Go backend, Postgres, Singapore Kubernetes, multi-tenant isolation with row-level security
  • JE Engine / JE Agent single-turn tool loop; TIDAR KAG sourced Q&A
  • Knowledge graph and mastery (compute-on-read); Wilson lower-bound judgement
  • Multi-agent grading with review; teacher voice cloning (CosyVoice); photo recognition (Qwen-VL)
  • Email verification codes, offline payment ledger, verifiable certificates, in-app assistant and tickets, seven-language interface packs

Built for this project

  • Cross-border weak-network live access with three-level degradation (meeting-tool deep links first)
  • Recording transcoding, multi-bitrate delivery, offline content packs and school cache points
  • Offline-first client: PWA first, native Android later; shared-device multi-account
  • Per-learner agent roster and four-layer memory (profile / mastery / episodic / working)
  • Chinese ASR and pronunciation assessment pipeline (Phase 2); HSK graph subtree and item bank (curriculum engineering)

Eight PoC cards (September–October, 4–5 weeks)

#What is validatedMethodPass criteriaIf it fails
PoC-1Cross-border live usability (per country)A China-side teacher delivers three 30-minute classes on Zoom and three on Tencent Meeting to each pilot school across morning, afternoon and evening slots; five listeners score; drops and degradation events recordedSubjective audio clarity ≥ 3.5 / 5; ≤ 1 drop per class, recovered within 30 s; three-level degradation triggered and successfulThat country moves to recording-first with one weekly audio Q&A
PoC-2Content delivery (per country)Download a 50 MB pack over school 4G; compare Cloudflare, Alibaba Cloud CDN and direct OSS, five runs eachMedian ≤ 8 minutes; resumable download success 100%Night-time pre-download + school cache point
PoC-3Low-end AndroidInstall the PWA on a 2 GB-RAM Android 9 device; 20 minutes of practiceCold start ≤ 5 s; no crash; storage ≤ 150 MB; first bundle ≤ 1.5 MB gzipCut animation and font packs; plain-text mode
PoC-4Offline practice and syncIn flight mode complete three practice sets and one mock quiz; reconnect after 24 hoursZero data loss; sync ≤ 60 s; conflicts resolved by server authority + event mergeRebuild the sync engine; MVP reduced to offline review cards
PoC-5Study-companion memory loopOne learner, five sessions; verify profile, mastery and episodic memory write and recall; 50 in-syllabus questionsSourced rate 100%; error rate ≤ 2% with every error handed off; session 5 correctly references weaknesses from sessions 1–4Restrict the agent to review and reminders; all Q&A to teachers
PoC-6Shared-device multi-accountThree learner accounts on one phone; offline data isolatedSwitch ≤ 3 s; no cross-contamination; PIN protection; separate local encryptionMVP one account per device; sharing in Phase 2
PoC-7Local collection ledgerConfirm future payment methods with the school; record and reconcile one mobile-wallet or school-collected paymentOne payment recorded and reconciled in the systemPilot entirely free; payments in Phase 2
PoC-8Compliance feasibility (three countries)Counsel issues applicability memos for the three national data laws, Singapore PDPA and China PIPL, plus trilingual consent formsMemos signed; data schedule complete; no “cannot do” conclusionAdjust data flows (e.g. teacher data stays in China) before MVP

Figures in the pass criteria are Phase 0 working thresholds, not external commitments; they are revised after measurement.

MVP technical scope (frozen end of October)

ModuleMVP (end of October)Phase 2 (January–March)
LiveZoom or Tencent Meeting deep links; platform handles scheduling, roll call, recording retrieval, region-based entrance suggestionCommercial RTC adapter (per PoC-1)
RecordingsTeacher uploads; multi-bitrate HLS; subtitles (Chinese / English); pack downloadTeacher-voice narration; French / Kiswahili subtitles
ClientInstallable web app: offline practice, review cards, recordings, multi-account switchingNative Android (beta), iOS
AgentsStudy companion + review; profile and mastery memory; sourced answers, hand-off when uncertainPronunciation coach, conversation, assessment, homeroom agents
AssessmentObjective auto-marking; photo homework recognition + grading + teacher reviewSpeaking assessment, mock-exam system, verifiable completion certificates
Teacher sideClass insight, review queue, alerts, AI proposal inbox (web)Homeroom agent drafts communication and scheduling
School / enterprise sideRoster import, attendance weekly report, device and consent dashboardTerm report for authorities, HR dashboard, bulk certificate verification
OperationsReused admin console: tenants, schools, classes, learners, content review, language packs, support ticketsChannel and partner management, settlement, external AgentOps board
Scaffolding languagesInterface in Chinese / English / Kiswahili / French; English Q&A scaffoldingFrench and Kiswahili explanation scaffolding; Arabic interface
Online libraryWeb reader for anthologies and Future Academy readers: shelf, progress, highlights, tap-to-defineTeacher-voice read-aloud, shared annotation, auto-listing of new titles, reading data into the graph

Architecture decision records (written within Phase 0)

  • ADR-1 Tenancy: GACEE is the top-level tenant, each country a region label, each school a sub-tenant; data isolated per school, cross-school aggregation only in the Association view; Postgres row-level security with automated isolation tests.
  • ADR-2 Live source: MVP uses meeting-tool deep links behind a media adapter; switching to commercial RTC changes no business code.
  • ADR-3 Offline strategy: PWA first, Service Worker + IndexedDB for packs and practice; the native app reuses the same sync protocol and pack format.
  • ADR-4 Model routing: all generation, recognition and synthesis go through the Model Gateway; Qwen family by default for Chinese-teaching Q&A, Claude available for review; no on-device models when offline.
  • ADR-5 Memory storage: profile and mastery structured; episodic memory append-only, exportable, deletable; memory is never treated as fact.
  • ADR-6 Data residency: core data in Singapore; teacher personal data minimised in China under PIPL; learner data leaves each country only under local law and guardian consent; no media node in China.
  • ADR-7 Cost guardrails: monthly model and traffic budgets per tenant; over budget degrades to recordings and local review; no silent overspend.
06 · Architecture & technical route

Architecture and technical route技术架构与技术路线

HanQiao's technical base is the JE Engine, the Agent Harness and TIDAR KAG already in production at JurisEdu. This project adds three layers: cross-border weak-network media (live / recordings / content packs), an offline-first client (PWA / app), and per-learner personalised agent orchestration. This chapter specifies component design, interface conventions and acceptance metrics; every latency, bandwidth and accuracy figure is a working assumption to be measured in Phase 0.

Architecture principles

Cloud-neutral (declarative Kubernetes deployment, S3-compatible storage, OpenTelemetry observability, Model Gateway abstracting model vendors); data resident in Singapore; agents propose and never execute writes; every learning artefact carries append-only hash-chain evidence; every network segment has a defined degradation path.

1. Geographic topology: three network segments

China · teacher access

  • Teacher client (web / desktop): live publishing, board annotation, review queue, class insight
  • Cross-border access via the meeting tool's mainland nodes or global acceleration; the platform carries only signalling and business data and deploys no media node in China (ADR-6)
  • Homeroom agent (Phase 2): scheduling, alerts, drafted communication

Core region · Agent Harness (Singapore)

  • Client gateway: ingress, OIDC authentication, per-tenant rate limits, media signalling
  • Media layer: meeting-tool / RTC adapter, recording retrieval, transcoding, VOD origin
  • Domain services: courses, scheduling, assessment, learners, certificates, collection ledger
  • JE Engine · JE Agent hybrid orchestration: Router → specialist agent → tools → Verifier → hand-off
  • Knowledge and memory: Chinese course graph, mastery, four-layer memory
  • Data / storage, observability, security, compliance, audit trail

Overseas · learner access (pilot schools)

  • Edge CDN / school cache point: recordings and packs served nearby
  • Learner PWA (offline-first): local store, packs, on-device practice and marking
  • Personal agent roster: companion and review first; pronunciation and conversation later
  • Shared-device multi-account: separately encrypted local store per account

Real-time and API traffic use the accelerated path; recordings and packs use the CDN; each segment is optimised separately, and degradation of any segment triggers the paths in section 6 of this chapter.

2. The ten-layer Agent Harness

The Harness packages the full agent lifecycle — build, run, evaluate, govern — as ten reusable layers; HanQiao is one vertical instance, and later language or subject programmes reuse the same runtime. Adding an agent requires only a definition manifest, tool bindings and an evaluation set, with no infrastructure change.

LayerComponent specificationUsed by HanQiao
UsersLearners · teachers · operations / academic office · schools · enterprises · local facilitators; isolated by tenant and schoolSix roles, one interface each (§07)
EntryPWA / app · web · live / recording clients · REST API · webhooks · workflow triggers; OIDC single sign-onMVP: PWA + meeting-tool deep links; roster-import API for schools
OrchestrationJE Agent hybrid orchestration: deterministic Router → specialist agents → Verifier → hand-off; task decomposition, tool selection, memory read/write, reflection, multi-agent collaboration, state and context management; agent manifest (YAML: role, tool allow-list, memory scopes, evaluation set, guardrail policy)Companion + review first; pronunciation / conversation / assessment / homeroom by phase
CapabilitiesBuilt-in tools: graph retrieval · ASR · pronunciation scoring (GOP) · TTS · item bank · spaced repetition (FSRS) · grading · handwriting recognition; external tools: object storage · CDN · meeting tools / RTC · messaging · collection ledger; the registry separates read-only from write toolsReuses multi-agent grading, Qwen-VL recognition, CosyVoice voice; builds ASR / GOP
ModelsModel Gateway: routing by slot (generate / review / recognise / synthesise); cost-, latency- and residency-aware selection; circuit breaking and failover; version canaries; response sanitisation. Model Zoo: JE Engine · Qwen / Model Studio · Claude · Gemini · open-source · privateGateway in production (R4 below)
DataChinese course graph · vector store · business database (Postgres, row-level security) · cache (Redis) · object storage (S3-compatible) · personal memory; governance: quality, lineage, access control, audit, de-identification, lifecycleGraph extended with the HSK system (section 4)
Evaluation and monitoringAgent evaluation (automated + human + business metrics) · end-to-end traces / spans (OpenTelemetry) · tokens and cost · sourced rate · error rate · guardrail hits · hand-off rate · alert boardsPilot KPIs read directly from this layer
Governance and securityIdentity and permissions (RBAC + tenant scope) · policy management · content safety · compliance audit · de-identification / encryption (AES-256 at rest, TLS 1.3 in transit) · cost and quotasTenant budget guardrails (ADR-7)
InfrastructureKubernetes · containers · service mesh · autoscaling; CI / CD with immutable images; multi-active, backup and disaster recovery, monitoring; cross-border acceleration, multi-CDNSingapore cluster and release model reused
Agent lifecycleDefine → build → test and evaluate → canary deploy → monitor → optimise; release gate = no regression on the three evaluation metricsEvery agent has an evaluation set and a canary channel

3. JE Engine and hybrid orchestration: six capabilities verified in production

The six capabilities below are live in JurisEdu production and verified end-to-end; HanQiao inherits them directly. Each is constrained to be grounded, auditable and human-decided.

#CapabilityImplementationMeaning for HanQiao
R1Intent / domain routingDeterministic heuristic routing (no model call, zero latency) sends a request to the learning / account / schedule domain, injects that domain's grounded context, then answers: route → ground → answerA learner's question is classified first; learning questions automatically carry the learner's calibrated weak points
R2Cross-capability tool registryRead-only tools join the coach's tool loop; write tools use native function-calling and always follow propose → approve → execute; approvals are written to an append-only ledger with SHA-256 chained hashes (previous hash + record)“The agent proposes practice, the teacher approves generation” is the technical form of teacher sign-off
R3Persistent learner memoryCross-session memory has only two write paths: observations derived from real mastery snapshots, and goals stated verbatim by the user; repeated observations accumulate salience; insufficient evidence is labelled “few samples”Memory accumulated per learner cannot be fabricated by construction
R4Model gatewayPlaced before the existing PII de-identification → Singapore residency → failover → metering chain; adds only slot policy and response sanitisation; any gateway error falls back to direct callsRoutes by cost and compliance without bypassing any guardrail
R5Per-run traceEach turn records route / ground / recall / tool / verify / end steps, storing safe metadata only, never question or answer text“Why this answer” can be replayed during the pilot; hand-off rate is measurable
R6Educational TIDAR KAGSection 4Core of low-hallucination Q&A and mastery judgement
Creation chain

Every learning artefact (homework, essay, assessment, certificate) carries an append-only hash-chain record. Models have no tool that writes to the chain; a failed write surfaces as an error rather than being silently ignored. The creation chain is the technical source of trust for HanQiao certificates and learning reports.

Orchestration loop (one turn)

StepActionBasisLatency budget (assumed)
1 InputLearner's item, question or speech; profile memory and mastery snapshot (as-of now) injected at session startR3
2 RouterDeterministic routing fixes domain and intent; relevant episodic memory recalled (top-k by salience and recency)R1 + R3≤ 50 ms
3 Specialist agentCompanion / review / pronunciation / conversation / assessment / homeroom calls allow-listed tools per manifest; Q&A retrieves through TIDAR KAG with sourcesR2 registry≤ 2.5 s (P50)
4 VerifierConsistency check; grading is written only when the scoring and review agents agree, otherwise the review verdict is taken and the item is downgraded to “for teacher”Multi-agent grading pattern≤ 1.5 s
5 Answer or hand-offSourced, personalised answer; low confidence, subjective items and write actions go to the teacher queueAbstention gate + propose → approve
6 Write memorySession summary to episodic memory; answer events update mastery; trace storedR3 + R5Asynchronous

4. TIDAR KAG: the source of low hallucination

TIDAR KAG is the platform's own knowledge-graph, provenance-bound generation, distinct from generic vector RAG. It was validated in the legal domain first, then ported to the educational course graph and is in production. Three pillars provide the technical guarantee against fabrication.

Temporal

Every judgement rests on a real answer snapshot as of a point in time; two points can be diffed per knowledge point — progress, regression, newly assessed, unchanged — and reported as such.

Provenance

Every conclusion is attached to real evidence: answer records, missed items, prerequisite and related neighbourhoods in the course graph; answers carry sources, and without a source there is no answer.

Calibrated

Mastery is not a point estimate but the lower bound of a Wilson confidence interval (z = 1.96); fewer than three attempts is “insufficient evidence” and the agent abstains; zero attempts is “not yet assessed”.

Course graph structure

Hierarchy
System → stage → subject → domain → topic → knowledge point; only structures listed by an authoritative standard are written
HanQiao subtree
“International Chinese HSK 3.0 (three bands, nine levels) / YCT” system: vocabulary, grammar, characters and scenes as four knowledge-point types, each with level, source, examples and prerequisite edges
Edge types
Prerequisite, related, scene-contains, confusable
Evidence nodes
Answer events (item ID, result, time, source client), grading verdicts (two-agent agreement), teacher sign-off records
Mastery
Compute-on-read: aggregate as-of snapshots per knowledge point, output the Wilson lower bound and sample size; conclusions are never cached, avoiding staleness
MechanismRuleWhere it lands in HanQiao
Knowledge boundaryQ&A retrieves and generates only within the course graph and textbook; out-of-scope requests return “outside this course, sent to your teacher”HSK Level 1 subtree + six-week textbook slice
Abstention gateLow confidence or subjective items abstain and go to a personAbstentions enter the teacher review queue; the weekly report counts the hand-off rate
Memory is not factMemory serves retrieval and personal style only; facts come from the graph and live masteryClass changes and re-learning correct automatically
Two-agent agreementScoring and review agents judge independently; disagreement takes the review verdict and downgrades to unknown; only clear right / wrong is written to the graphPhoto homework and objective items; speaking disagreements go to the teacher
Write separationAgents only propose; a person approves and the system executes with a hash-chain recordScheduling practice, adjusting progress and issuing certificates all go through approval
Evaluation setEvery release runs the in-syllabus set: sourced rate, error rate and hand-off rate must not regress50 items in Phase 0, 200 for MVP, 500 in Phase 2

5. Per-learner agent roster and four-layer memory

Study Companion

Plans the path, adjusts difficulty, explains knowledge points; Q&A via TIDAR KAG with sources; runs pre-packaged tasks when offline.

Review Agent

Schedules spaced-repetition cards by the FSRS forgetting curve and mastery; scheduling runs on device and works offline.

Pronunciation Coach (Phase 2)

ASR + tone / phoneme scoring (GOP), character-level correction, demonstrations in the teacher's voice.

Conversation Agent (Phase 2)

Role play — ordering, asking directions, interviews — with teacher-voice TTS and gentle correction.

Assessment Agent (Phase 2)

Assembles quizzes and mock exams; multi-agent grading and review; speaking items via the pronunciation pipeline; subjective items abstain to the teacher.

Homeroom Agent (teacher side, Phase 2)

Scheduling, progress and risk alerts, make-up suggestions, drafted parent communication; every action goes through proposal and approval.

Memory layerContentStorageRead / writeData-subject rights
ProfileTarget level, mother tongue / scaffolding language, time zone, device and network profile, correction preferences, available slotsStructured (Postgres)Injected at session start; updated on settings changeViewable, editable
Semantic / masteryMastery and error patterns per knowledge point, word and grammar itemKnowledge graph (compute-on-read)Updated by events after each answer and assessmentExportable
EpisodicSession summaries, error causes, milestones, motivation signalsAppend-only event log + vector storeWritten at session end; recalled by relevanceExportable, deletable
WorkingCurrent session contextIn-session (Redis, TTL)Valid within the session; distilled to episodic memory at closeNot persisted

6. Cross-border weak-network live, recording delivery and offline client

Live optionKey pointsTrade-off
A Existing meeting toolsZoom / Tencent Meeting carry the class; the platform integrates by deep link and handles scheduling, roll call, recording retrieval and classroom-event capture; entrance suggested by learner regionLowest cost, fastest; limited classroom customisation. Adopted for MVP
B Commercial RTC SDKFor custom classroom / whiteboard / deep embedding, an SDK with a global real-time network (Alibaba Cloud RTC / Agora / Tencent RTC / ZEGO)Controllable experience; per-minute billing. Evaluated in Phase 2 per PoC-1
C Self-hosted SFUWebRTC SFU (LiveKit / mediasoup) + own TURN + cross-border accelerationLow marginal cost, heavy operations; evaluated at scale

Three-level degradation and recording fallback

  • Level 1: fully interactive live (audio-video + whiteboard + live Q&A)
  • Level 2: camera off; audio + board / slides + captions + text chat retained
  • Level 3: network unavailable → recording available within 15 minutes + agent-led practice to fill the gap
  • Every live class is cloud-recorded to VOD by default; dropped learners count as replay attendance; degradation events flow to the classroom event log

Weak-network engineering checklist

  • Adaptive bitrate + simulcast / SVC; resolution and frame rate degrade automatically; audio and whiteboard prioritised
  • Opus + DTX + FEC / NACK / PLC against loss; adaptive jitter buffer
  • TURN relay when NAT traversal fails
  • Audio-first mode: voice + board + captions only

Recordings and content delivery

  • Teacher recording or live capture → origin object storage; automatic subtitles (Chinese / English / French); teacher-voice narration in Phase 2
  • FFmpeg HLS ladder: 240p 300 kbps · 360p 600 kbps · 480p 1 Mbps · 720p 2 Mbps · audio-only 48 kbps; 6-second segments; AES-128 per-pack keys
  • Origin → multi-CDN; African edges serve nearby with resumable download; the school cache point is a reverse cache on an always-on device
  • Content packs per lesson / unit (manifest + version + checksum); Wi-Fi-only download option; revocable on withdrawal

Offline-first client (PWA)

  • Available locally: downloaded recordings, in-class practice, review cards, stroke order, objective answering and marking, recorded audio
  • Local store (IndexedDB), versioned packs; outbound event queue with idempotency keys; retry with exponential back-off; resumable transfers
  • Sync protocol: event-sourced, server monotonic sequence as authority; progress = server authority + event merge; notes = last-write-wins
  • Storage budget ≤ 150 MB with tiered clean-up; data-saver mode; shared-device multi-account with per-account local encryption keyed from the PIN (Argon2id)

7. Assessment pipeline, model layer and cloud-neutral matrix

Listening · reading · writing · character writing

Objective items auto-marked; photo homework recognised by Qwen-VL → two-agent point-by-point grading → teacher review and sign-off; handwriting recognition checks stroke order and structure; every answer updates the graph by event.

Speaking and pronunciation (Phase 2, selection to be validated)

Chinese ASR (Paraformer / SenseVoice class) transcription; forced alignment (CTC) + GOP goodness-of-pronunciation scoring; tone classified from the F0 contour (pYIN); corrective demonstrations in the teacher's voice; disputed items go to the teacher.

DependencyPrimary (current baseline)AlternativesPortable abstraction
Compute / orchestrationAlibaba Cloud ACK (Kubernetes, Singapore)AWS EKS · GCP GKE · Azure AKS · self-hostedDeclarative Kubernetes manifests
Object storageAlibaba Cloud OSSAWS S3 · GCS · MinIOS3-compatible API
Database / cache / vectorPostgres · Redis · vector storeManaged or self-hosted equivalentsOpen-source components, no proprietary dependency
Live / RTCExisting meeting toolsAgora · Tencent RTC · ZEGO · self-hosted SFUMedia adapter
CDNMulti-CDN orchestrationCloudflare · Akamai · CloudFront · Alibaba Cloud CDNChosen per country by measurement
Cross-border accelerationCloud global accelerationOther accelerators / leased linesSwitchable access layer
LLM / ASR / TTSModel Studio (Qwen · Paraformer · CosyVoice)Claude · Gemini · open-source · privateModel Gateway
Observability / securityLogs · metrics · traces · WAF · key managementOpenTelemetry + open-source stackOTel standard

Access and content security

Token authentication, end-to-end TLS, tenant isolation, least privilege; encrypted recordings, encrypted and revocable offline packs, optional watermarking.

Assessment integrity

Randomised item banks, time limits, tab-switch and multi-device detection, live read-aloud for speaking; in-person centres recommended for high-stakes exams.

Agent guardrails

Protection against privilege escalation and prompt injection, output review, memory-write validation, audit trail against memory poisoning; multi-active media and CDN degradation.

8. Non-functional targets (pilot acceptance baseline, revised after measurement)

DimensionMetricTarget (assumed)Measurement
Low-end devicesCold start · first bundle · storage≤ 5 s · ≤ 1.5 MB gzip · ≤ 150 MB2 GB-RAM Android baseline
OfflineSync success · data loss≥ 99.5% · 0Client telemetry + server reconciliation
AgentsSourced rate · error rate · hand-off rate100% · ≤ 2% · recordedEvaluation set + 50 sampled items reviewed weekly
LiveAudio clarity · degradation triggers · replay availability≥ 3.5 / 5 · recorded · ≤ 15 minPoC-1 method carried into the pilot
AvailabilityCore-region monthly availabilityRecorded during the pilot, not committed; 99.5% in formal operationObservability layer
SecurityTenant isolation tests · penetration testEvery release · once before MVP freezeCI gate + external test
CostModel + traffic + meeting tools per learner per monthMeasured model with per-tenant budget guardrailsMonthly cost table

9. Technical roadmap (aligned with milestones)

PhaseTechnical goalHarness incrementWindow
Phase 0 · PoCCross-border weak-network measurement in three countries, CDN coverage, low-end devices, offline sync, memory loop, multi-account, collection ledger, complianceTenancy model (ADR-1), media adapter (ADR-2), companion evaluation set of 50September–October, 4–5 weeks
Phase 1 · MVPRecordings + meeting-tool live + PWA offline + companion and review agents + objective items and photo homework + teacher review queue + school weekly reportHSK Level 1 graph subtree, profile and mastery memory, tenant budget guardrails (ADR-7)Mid-September to end of October, 6-week build; pilot from November
Phase 2 · EnhancementASR / GOP pronunciation, conversation, assessment and homeroom agents, verifiable completion certificates, native Android, French / Kiswahili scaffolding, corporate classesEpisodic memory with vector recall, write approval extended to scheduling and certification, evaluation set 200–500January–March, about 12 weeks
Phase 3 · ScaleMulti-country multi-school, commercial RTC or hybrid, multi-CDN cost optimisation, authority reporting and certification alignmentMulti-active deployment, external AgentOps board, unit-economics-driven gateway routingFrom the second quarter of next year

Key assumptions to verify

#AssumptionImpactWhen
1Cross-border real-time quality from China to the three countries (latency, loss, time of day)Live-first or recording-first per countryPhase 0
2Cost, compliance and lock-in of commercial RTC versus self-hosted SFUCost structure and engineering loadPhase 0 → 1
3Actual CDN presence and unit price in the three countriesRecording experience and traffic costPhase 0
4Whether a media or access node in China is lawful and requiredLegality of the cross-border pathDecided up front
5Accuracy of Chinese ASR and pronunciation modelsCredibility of speaking assessmentPhase 1 → 2
6Feasibility of on-device pronunciation scoring on low-end AndroidScope of offline speaking practicePhase 1
7Curriculum production capacity (courses, item bank, HSK alignment)Teachable levels and paceThroughout
8Path to official certification and issuanceCredibility of outcomesPhase 2
07 · Applications & scenarios

Applications and scenarios应用与场景

This chapter describes the interface used by each of six roles and the twelve most frequent scenarios in the pilot. Each scenario lists trigger, participants, system steps, components, success criteria and exception handling, and serves as the common basis for the requirements document, test cases and training. The demonstration environment at learn.gacee.org implements interactive prototypes of all interfaces described here.

1. Interfaces by role

RoleEntryCore screensMVPPhase 2
LearnerPWA (phone / shared device) · learner ID or phone + PINToday, courses and recordings, practice and review cards, photo homework, live entry, library, knowledge map, my agents, certificatesAll except pronunciation and conversationPronunciation coach, conversation, native app
Remote teacherWeb (desktop) · email codeClass insight, review queue, AI proposal inbox, live and recording, scheduling, class library, weekly reportAll except homeroom agentHomeroom agent, voice narration
School (principal / office / facilitator)Web or mobile browser · email codeRoster import and class set-up, learner and device dashboard, consent coverage, attendance weekly report, term reportRoster, dashboards, weekly reportAuthority term report, bulk certificate verification
Enterprise (HR)Web · email codeCommissioned groups, attendance and progress reports, completion certificates, invoicesAll
Operations (Association office / platform)Admin console · two-factorTenants and schools, content and publishing review, AgentOps, compliance register, cost and budget, network and delivery, language packs, support ticketsAllChannels and settlement
Public / verifiersverify.gacee.orgVerify by certificate number or QR; minimum information onlyAll

2. Twelve delivery scenarios

S1 · A learner's week (online / weak network / offline)

Trigger
Monday release of recording and practice pack; the Study Companion builds the week's tasks from the learner profile (available slots, power-cut forecast)
Participants
Learner; Study Companion and Review agents; teacher (passive)
Steps
Online: watch recording → in-class practice → review cards. Weak network: audio-first mode + text practice. Offline: local pack + review cards + on-device marking; outbound queue syncs on reconnection
Components
Content pack, PWA local store, Review Agent (on-device FSRS), sync protocol, mastery events
Success
The 15-minute daily task completes on all three paths; zero sync loss; weekly mastery gain visible
Exceptions
Low storage → tiered clean-up prompt; pack checksum failure → re-download the affected segment

S2 · A cross-border live class (teacher in China, learners in three countries)

Trigger
Scheduled time; the platform suggests Zoom or Tencent Meeting by region and sends trilingual reminders
Participants
Lead teacher, local facilitator, learners, operations on duty
Steps
Teacher starts from the teacher client → platform creates the meeting and captures its ID → learners join by deep link with entry time recorded → level 1 / 2 degradation triggered by network quality → recording retrieved and transcoded after class → replay within 15 minutes → absent learners counted as replay attendance
Components
Meeting-tool adapter, roll call and classroom event log, recording retrieval, transcoding, notifications
Success
Audio clarity ≥ 3.5; degradation events recorded; replay within 15 minutes; attendance (incl. replay) computed automatically
Exceptions
Meeting tool unavailable → backup link on the other tool; teacher drops → facilitator hosts + recording substitutes (level 3)

S3 · Offline practice during a power cut and next-day sync

Trigger
Power-cut window in the profile memory (e.g. 19:00–21:00); the Companion packs the day's tasks for offline use and moves the reminder to 17:30
Steps
Learner completes review cards and watches the recording offline → events queued → synced in sequence next day → mastery updated → teacher client shows “completed offline”
Success
PoC-4 criteria; the teacher client distinguishes online from offline completion
Exceptions
Device clock drift → server sequence prevails; duplicate submission → de-duplicated by idempotency key

S4 · Photo homework to teacher sign-off

Steps
Learner photographs and uploads (stored locally first when offline) → Qwen-VL recognises handwriting and layout → scoring and review agents grade independently → agreed items written to mastery; disagreements and subjective items enter the teacher queue → teacher confirms or amends item by item → sign-off → result and comments return to the learner; the whole process is written to the creation chain
Success
Two-agent agreement rate recorded; the teacher sees only disagreements and subjective items; zero undetected errors
Exceptions
Blurred photo → prompt to retake with an example; low recognition confidence → whole submission to the teacher

S5 · Agent Q&A and hand-off

Steps
Learner asks → deterministic routing classifies as learning → weak points injected → TIDAR KAG retrieves within the course graph → sourced answer with citation; low confidence or out of scope → abstain, question enters the teacher queue, answered in Wednesday's live class or in writing; trace stored
Success
Sourced rate 100%; handed-off questions closed with a record; 50 items reviewed by a person weekly

S6 · Three learners on one shared device

Steps
Home screen shows “Who is learning?” → choose avatar and enter PIN → that account's encrypted local store unlocks → study → sign-out locks; packs shared by reference count, practice data isolated
Success
Switch ≤ 3 s; no cross-contamination; forgotten PIN reset after school or parent confirmation

S7 · A teacher's week

Steps
Monday: publish recording and practice pack (reusing past recordings) → Wednesday: 30-minute live class → daily: alerts (absence, mastery regression, power-cut impact) → Friday: clear the review queue (disagreements and subjective items only) → approve AI proposal inbox (practice, progress, make-up) → weekly report generated automatically and signed off
Success
≤ 6.5 teacher hours per class per week; one teacher serves three schools; every write action has an approval record

S8 · School side: roster to term report

Steps
Office uploads the roster (CSV or photo) → agent proposes classes (never invents names) → school approves → accounts and PINs issued → consent dashboard → device and cache-point status → weekly attendance (incl. replay) and practice report → term report for the education authority (English / French)
Success
100% consent coverage before opening; weekly report delivered to principal and facilitator; report exportable

S9 · Corporate commissioned class (Phase 2)

Steps
HR submits roster and goal (workplace scenarios) → class opens → HR dashboard shows attendance, progress and completion only, never answer detail → certificates issued in bulk → settled per class with invoice
Success
Employees hold the same data-subject rights as school learners; HR access limited to report fields

S10 · Library: from competition essay to shared reading

Steps
Winning essays → publisher editing → level tagging → auto-listed in the library → teacher assigns shared reading → learners highlight, tap to define, annotate → new words to review cards → reading data into the knowledge graph → strong class work nominated for the next anthology
Components
Publishing workflow, reader (whole book cached offline), tap-to-define, shared annotation, graph events
Success
Weekly reading time and word-to-card conversion measurable; new titles listed automatically

S11 · Certificate issuance and verification (Phase 2)

Steps
Course ends → system proposes issuance (attendance ≥ 80% and mock exam passed) → teacher signs (creation chain) → Association seal and number → learner receives PDF and QR → a verifier enters the number at verify.gacee.org and sees validity, initials, course and level only
Success
Verification ≤ 2 s; revocation immediate; bulk verification by CSV

S12 · Support and tickets (multi-channel, agent-assisted)

Steps
WhatsApp / in-app assistant / email / school group / phone → one ticket inbox → classify (downloads, account, reports, schedule, data request) → ground (delivery logs, runbooks, policy) → recall same-school history → allow-listed low-risk actions executed automatically and logged (e.g. push the audio pack) → high-risk actions (identity, deletion) to a person → drafted reply; auto-sent when confidence ≥ 90% and allow-listed, otherwise held for review
Success
Median first response ≤ 5 minutes; auto-resolution rate recorded; data requests closed within 48 hours

3. Interface and scaffolding languages

Interface packs cover Chinese, English, Kiswahili, French, Spanish, Dutch and Indonesian, managed centrally by operations (set, review, correct, publish); unreviewed entries never reach production. Scaffolding (explanations) is English in the MVP, with French and Kiswahili added in Phase 2; the three pilot schools default to English, Kiswahili and French interfaces respectively.

08 · Workstream D

Delivery落地实施

This chapter gives the month-based master timeline, work breakdown structure, team, responsibility assignment, budget structure, risk register, quality acceptance and pilot operations manual. All windows are relative to the start month; if approval slips, every window moves with it.

Master timeline (seven months from the start month)

Work itemSepOctNovDecJanFebMar
A · Approval, school shortlist, LOIs and teacher agreements (4 weeks)
A · Pilot MOU and data schedule (3 weeks)
A · Education authorities, enterprises, channels (rolling)
B · Needs research and field measurement (2–3 weeks per country)
C · Phase 0 PoC (8 cards, 4–5 weeks)
D · MVP build and freeze (6 weeks)
D · Pilot preparation (2 weeks per school)
D · Three-country pilot (6 weeks per school, staggered ≤ 2 weeks)
D · Review and expansion decision (2 weeks)
D · Phase 2: speaking, conversation, certificates, corporate (about 12 weeks)
D · Native Android app (10–12 weeks)
Curriculum: HSK Level 1 course and item bank (continuous)
Compliance: registers, data schedules, consent (front-loaded, then continuous)
Milestones
MilestoneWindowGate (no pass without)
M1 PoC gateEarly October (week 5)Eight cards concluded; live / recording emphasis decided per country; ADR-1 to 7 written; no “cannot do” in the three compliance memos
M2 MVP freezeEnd of October (week 8)Scope locked; three teachers trained; devices and consent in place at all three schools; first six weeks of HSK Level 1 content reviewed
M3 Pilot reviewLate December (within 2 weeks of the last school closing)Six success criteria in §00 each evidenced; unit-cost model measured; cross-country conclusions; expansion list confirmed
M4 Phase 2 acceptanceMarchSpeaking assessment and conversation agent live; verifiable completion certificates issued; first corporate class or paid seats in place

Work breakdown structure

#Work packageMain tasksOwnerWindow / duration
1.0InitiationBoard resolution, joint programme office, tools and document library, fact-sheet regimeAssociation lead + platform leadSeptember, 2 weeks
2.0External engagementSchool shortlist, LOIs, pilot MOU and data schedule, teacher agreements, consent localisationRay · Liu Laoshi · counselSeptember–October, first round 4 weeks
3.0Needs researchInterviews, questionnaires, field measurement, content audit, PRD, personas, journeys in three countriesAssociation + platform product; facilitators executeSeptember–October, 2–3 weeks per country
4.0Technical validationEight PoCs, ADRs, cost model, MVP scopePlatform engineeringSeptember–October, 4–5 weeks
5.0CurriculumHSK Level 1 syllabus, six weeks of courseware and recordings, 300 items, 400 review cards, review processLiu Laoshi + China-side teachersFrom September, 6 weeks before MVP; continuous
6.0MVP buildTenants and classes, recording pipeline, PWA offline, companion and review agents, photo homework, teacher, school and operations clientsPlatformMid-September to end of October, 6 weeks
7.0Pilot preparationSchool devices and network, cache point, consent collection, account issuance, teacher and facilitator training, support channelsRay + schools; teacher training by Liu Laoshi2 weeks before each school opens
8.0Pilot operationsSix weeks per school, weekly reports, issue tracking, data collection, follow-up interviewsJoint programme officeNovember–December
9.0Review and expansionReview report, unit-economics model, cross-country comparison, expansion list, delivery handbook v1Association lead + platform leadDecember, 2 weeks
10.0Phase 2 buildSpeaking assessment, conversation agent, certificates, native app, scaffolding languages, corporate classesPlatform + AssociationJanuary–March, about 12 weeks
11.0Compliance and securityThree-country registers, data schedules, consent, content review, audit trail, penetration testCounsel + platform securityThroughout

Team

RoleHeadcountSourceCommitmentDuties
Programme sponsor1Association lead3 hours per weekDecisions, resources, external relations and contracting
Platform lead (technology)1Platform1 day per weekTechnical route, architecture decisions, team allocation
Programme manager1Platform (or Association appointee)Full timePlan, progress, risk, weekly meetings, documents, contract register
Pedagogy lead1Liu LaoshiHalf timeCurriculum, resource packs, final item review, teacher recruitment and coordination
Africa business lead1RayHalf timeCandidate schools and local partners in three countries, authorities, payment channels, market partners
Resource advisers2Zhang Laoshi, Publisher LiAs neededCurriculum review and expert guidance; publishing and media
Chinese teachers (China side)2–3Engaged by the Association6–10 hours per week + preparationRecording, live classes, review, sign-off
Operations / academic office1AssociationHalf timeScheduling, learner support, parent communication, data registration
Backend engineers2PlatformFull time to MVP freeze, then 1Tenancy, courses, recordings, sync, agent wiring
Front-end / mobile engineers1–2PlatformFull timePWA, teacher and school clients; native app after the pilot opens
AI / agent engineer1PlatformFull timeMemory, orchestration, evaluation sets, gateway routing
Design1PlatformHalf timeBrand application, learner interface
QA / operations1Platform (shared)One thirdWeak-network and low-end testing, releases, monitoring
Legal counsel1External (local counsel in three countries as needed)About 8 days in Phase 0, then as neededThree-country memos, data schedules, consent, contract localisation
Local facilitator1 per schoolPilot school4 hours per weekDevices, roll call, on-site support, measurement

Responsibility assignment (RACI)

ItemAssociation leadPlatform leadProgramme managerLiu Laoshi (pedagogy)Ray (Africa)Platform engineeringPilot schoolCounsel
Approval and budgetACRCCIIC
School selection and LOIsAICCRICC
Pilot MOU and data scheduleACRIRCCR
PoC conclusions and MVP scopeCARCCRIC
Curriculum, packs and teachersICIA/RICCI
Data and compliance designARCICRCR
Pilot operations and weekly reportsICA/RRRCRI
Authorities and market partnersAICIRICC
Review and expansion decisionARRCRCCI

A accountable · R responsible · C consulted · I informed

Budget structure (cost lines and estimation method; no amounts assumed)

Cost lineDriverEstimationCost lever
Association staffTeacher hours, curriculum, operationsHourly rate × hours + half-time salariesRecording reuse; one teacher serves three schools
Platform staffPerson-months (table above)Recorded per person-month as programme inputReuse existing capabilities; strict MVP scope
Meeting toolsHost seats2–3 paid seats per monthExisting tools first; no RTC minutes purchased
CDN and storagePack size × learners × downloads × countriesModelled after PoC-2 unit prices in three countriesLow bitrate, school cache points, night downloads
Model callsSessions per learner per week × tokens per sessionModelled after PoC-5; per-tenant budgetGateway cost routing; cache frequent answers
LegalCountries × instrumentsPer day; one local counsel per countryTemplated consent and data schedules
School supportDevice and data subsidies, facilitator stipends, cache-point devicePer school per monthShared devices, offline packs
TravelSchool visits, opening daysPer trip; video interviews preferredOne trip by Ray covers several schools

The review produces a measured cost per learner per month for each country, the sole basis for later pricing discussions with education authorities, employers and foundations.

Risk register

#RiskLikelihoodImpactResponseOwner
R1Cross-border live quality below threshold (one or all countries)HighHighArchitecture is recording-first; PoC-1 decides emphasis per country; live keeps audio onlyPlatform lead
R2Insufficient curriculum capacity; nothing to teach at MVPHighHighStart in September with the first six weeks of HSK Level 1; reuse existing courseware; AI-assisted generation with teacher reviewLiu Laoshi
R3MVP timeline too tight; quality suffersHighMediumScope frozen per this plan; native app deferred; weekly demonstrable incrementsProgramme manager
R4A candidate school in one country cannot sign on timeMediumMediumOne alternate per country; staggered opening ≤ 2 weeks; worst case two countries proceed firstRay
R5Inadequate devices or power at a schoolMediumHighResearch first; shared devices and multi-account; offline packs; cache pointRay
R6Cross-border data or minors' compliance not closedMediumHighPoC-8 up front; trilingual consent; data minimisation; no class without complianceCounsel
R7Undetected agent errors damage trustMediumHighSourced answers + hand-off when uncertain; 50 items reviewed weekly; errors recorded openlyAI engineering
R8Teacher resistance or attritionMediumMediumTeachers co-design; tools reduce load rather than monitor; hours and voice licences are paidLiu Laoshi
R9Model and traffic costs run awayMediumMediumTenant budgets and degradation; monthly cost reportAI engineering
R10“HanQiao” name confused with the “Chinese Bridge” competitionMediumMediumHanQiao remains an internal code name; the external education brand is named separately after trademark searchAssociation lead
R11External figures cannot be verifiedMediumMediumFact-sheet regime; unverified figures never publishedAssociation content lead
R12Single-vendor dependency for cloud or modelsLowMediumCloud-neutral matrix and Model Gateway; every third party has an alternativePlatform lead
R13Local legal review delays signing in three countriesMediumMediumNon-binding LOI first; MOU with a common English master text and country schedulesCounsel

Quality and acceptance

  • Weekly demonstration: Friday, 30 minutes, real learner accounts under simulated weak network (256 kbps, 5% loss), showing the week's increment.
  • Low-end baseline: two 2 GB-RAM Android phones on the team; every learner feature accepted on them.
  • Agent evaluation set: 200 in-syllabus items run at every release; sourced rate, error rate and hand-off rate must not regress.
  • Security: one penetration test and dependency scan before MVP freeze; automated tenant-isolation tests.
  • User acceptance: signed acceptance sheets from teacher, learners and school, on whether a week of teaching and learning can be completed independently.

Pilot operations manual (per school)

Two weeks before opening

  • Roster and grouping (by age and level); 100% consent collected
  • Accounts and PINs issued; devices installed and packs pre-downloaded; cache point deployed
  • Two hours of training each for teacher and facilitator, including a degradation drill
  • Support channels: WhatsApp group + in-app assistant + tickets; duty roster

Weekly rhythm

  • Monday: recording and practice pack released
  • Wednesday: 30-minute live (or audio) class; absentees watch the replay
  • Daily: 15 minutes of agent-led practice and review cards
  • Friday: teacher clears the review queue; weekly report (attendance, homework, accuracy, cost, incidents)
  • Every two weeks: learner and parent follow-up

Key metrics (pilot dashboard)

DimensionMetricSource
EngagementWeekly active rate, attendance (incl. replay), median weekly practice minutesSystem logs
LearningWeekly mastery gain, mock-quiz pass rate, homework submission rateKnowledge graph and assessment records
Teacher leverageTeacher hours per class per week, review-queue handling time, hand-off rate, learners served per teacher hourTeacher client logs + work logs
TrustSampled agent accuracy, sourced rate, exposures before an error is caughtWeekly human review
ReachOffline practice share, sync success, recording completion, degradation triggers (per country)Client telemetry
CostPer learner per month: model, traffic, meeting tools, staff allocation (per country)Monthly cost table
ComplianceConsent coverage, data-request response time, security incidentsRegister
09 · MOU & contracts

Memoranda and contracts备忘录与合同

This chapter lists every external legal instrument required for delivery, the key clauses of each, and reference samples. Samples use an English master text with Chinese as the working translation, for use by the Association's counsel and local lawyers during localisation; this chapter is not legal advice. Every signatory is marked “to be confirmed” until execution.

1. Instruments and signing sequence

#InstrumentParties (TBC)NatureTimingLanguage
C1Letter of IntentAssociation ↔ pilot schoolNon-binding (except confidentiality)September, after shortlistEnglish / French
C2Pilot MOUAssociation ↔ pilot school (and governing body where applicable)Binding: confidentiality, data protection, IP, termination; otherwise intentOctober, before openingEnglish master + local translation
C3Data Protection ScheduleSigned with C2; the platform provides a processor undertakingBindingWith C2English
C4Guardian / learner consent formGuardian (minor) or learner (adult) → AssociationBindingBefore opening, 100% coverageEnglish / French / Kiswahili
C5Teacher Services AgreementAssociation ↔ individual teacherBindingFrom SeptemberChinese
C6Voice LicenceTeacher → Association (platform as technical executor)Binding, separate instrument, separately revocableWith C5 or Phase 2Chinese
C7Local Facilitator AgreementAssociation or school ↔ facilitatorBindingBefore openingEnglish / French
C8Corporate Programme AgreementAssociation ↔ enterpriseBindingPhase 2English / Chinese
C9Framework of CooperationAssociation ↔ education authorityMainly intentAfter pilot dataEnglish / French

2. Contract management process

  1. Templates: the samples here are the common templates; the Association's counsel finalises the English master text.
  2. Localisation: one local lawyer per country reviews data-protection, minors, labour and tax clauses and produces a localisation variance table.
  3. Signatory confirmation: Ray confirms the school's legal entity and authorised signatory; the Association lead confirms the Association's signatory.
  4. Execution: electronic signature by default (DocuSign or equivalent); where paper is required locally, scans govern and originals are retained.
  5. Register: the programme manager maintains the contract register (instrument, version, signing date, term, renewal reminder, data-schedule version).
  6. Changes: changes to scope, term or data purposes are made by supplementary agreement and mirrored in the compliance register.

3. Key clauses and reference samples

C2 · Pilot MOU (structure and sample clauses)

1 Purpose and scope
A six-week Chinese-learning pilot for 20–40 learners at [School]; content is the first six weeks of HSK Level 1; delivered through the learn.gacee.org platform.
2 Term
From signature until completion of the pilot review; either party may terminate on 30 days' written notice; termination does not affect Clauses 7, 8 and 10.
3 Association obligations
Provide course content, remote teachers, platform access, training and support; charge no fee during the pilot; issue learning reports and completion certificates.
4 School obligations
Organise learners, provide a fixed timetable and venue, appoint a facilitator, assist with consent collection, permit on-site network measurement.
5 Fees
Free during the pilot; the School bears no system cost; the Association charges learners and guardians nothing.
6 Intellectual property
Course content, materials and brand belong to the Association; learner work belongs to the learner and may be used by the Association for teaching and non-commercial display with consent; platform software is owned by the Association's technology service provider and licensed to the School for the pilot.
7 Data protection
Per Schedule C3; the Association is controller, the School joint controller of its own learners' data; cross-border transfer only with guardian consent and in compliance with local law.
8 Confidentiality
Both parties keep unpublished pilot data, content and commercial information confidential for two years after termination.
9 Publicity
Prior written consent for any public reference to the pilot; separate consent for learner images and work.
10 Liability and disputes
Except Clauses 6–9, this MOU creates no binding obligations; disputes go to negotiation, then [SIAC / local court, to be fixed by counsel].
11 After the pilot
The parties may enter a formal agreement after the review; the School may request export or deletion of its learners' data.
Sample clause · Purpose and non-binding nature

This Memorandum of Understanding records the intention of the Global Association of Cultural and Educational Exchange (“the Association”) and [School] (“the School”) to conduct a six-week Chinese-language learning pilot for 20 to 40 learners using the Association's learning platform. Except for Clauses 6 (Intellectual Property), 7 (Data Protection), 8 (Confidentiality) and 9 (Publicity), this Memorandum does not create legally binding obligations. No fee is payable by the School, the learners or their guardians during the pilot.

C3 · Data Protection Schedule (key points)

ClauseContent
RolesThe Association is controller; the School is joint controller of its own learners' data; the technology platform is processor acting only on the Association's written instructions
Data itemsLearners: name, age band, school and class, learning records, homework and audio; guardians: name and contact; facilitators: name and contact. No ID numbers, precise addresses or biometrics
PurposesTeaching delivery, learning assessment, reports and certificates, pilot research (de-identified aggregates); no advertising; no learner data used to train any general model during the pilot
Storage and transferCore data in Singapore; encrypted in transit and at rest; cross-border transfer under a lawful route in local law and guardian consent
RetentionDeleted 12 months after the pilot or 30 days after account closure, whichever is earlier; the School may request earlier export and deletion
Sub-processorsListed (cloud, CDN, meeting tools, model services); 30 days' notice of changes
Data-subject rightsAccess, rectification, export, deletion, withdrawal of consent; acknowledged within 48 hours, completed within 30 days
Security incidentsNotification to the School and regulator within 72 hours of discovery, as required locally
AuditThe School may request an annual summary of isolation and security testing

C4 · Guardian / learner consent (key points, trilingual)

  • Data items and purposes; storage in Singapore; explanation and legal basis of cross-border transfer.
  • Live classes are recorded and speech practice is stored for teaching and assessment; never for advertising or training general models.
  • How to withdraw consent and request deletion at any time; contact and response times.
  • A separate, unticked option for use of photographs, recordings and work in the Association's publicity or publications.
  • Adult learners sign for themselves; guardians sign for minors; the school keeps a copy.
Sample clause · Consent to recording

I understand that live classes are recorded and that my child's voice recordings from pronunciation practice are stored securely in Singapore and used only for teaching and assessment. I may withdraw this consent at any time by contacting [email], after which the recordings will be deleted within 30 days. ☐ I additionally agree that my child's photographs or written work may be used in the Association's non-commercial publications (optional).

C5 · Teacher Services Agreement (key points)

ClauseContent
ServicesRecording, live classes, review queue, sign-off of key judgements, curriculum participation; weekly hour cap and preparation hours
RemunerationHourly rate + one-off recording fee + per-item review; monthly settlement; tax per the teacher's jurisdiction
Recording copyrightCopyright in recordings vests in the Association with attribution to the teacher; playable to learners within the platform, not sold externally; after termination, recordings may continue to serve classes already opened until they close
Pedagogical judgementThe teacher has final authority over grading, sign-off and certificates; platform tools only advise
Personal dataTeacher data minimised; storage and transfer under China's PIPL; contact details never published
Confidentiality and conductLearner data confidentiality; child-safeguarding code of conduct; no private charging
Termination30 days' notice by either party; hand-over of classes in progress

C6 · Voice Licence (key points)

  • Scope limited to course narration, pronunciation demonstrations, library read-aloud and correction feedback; no other use.
  • Revocable at any time; generation stops and the voice model is deleted within 30 days; generated course audio is deleted once classes already opened close.
  • Separate consideration; the voice model is stored only in Singapore and never shared externally.

C8 · Corporate Programme Agreement (key points)

ClauseContent
ServicesClass format (size, weeks, target scenarios), delivery (recordings + live + agent practice), HR report frequency, completion certificates
Fees and settlementPer-class pricing (TBD); 50% before opening, balance on completion; invoiced by the Association
DataEmployee learners hold the same data-subject rights as school learners; HR sees attendance, progress and completion only; no answer-level detail
Service levelBest-effort during the pilot with availability recorded; 99.5% monthly availability and 24-hour response in formal operation
Intellectual propertyCourse content belongs to the Association; enterprise-specific content (glossaries, scenarios) belongs to the enterprise with a teaching licence to the Association
Termination30 days' notice; classes in progress settled pro rata

C9 · Framework of Cooperation with an education authority (key points)

  • Scope: curriculum alignment, expansion to further schools, certification pathway, teacher training.
  • Data sovereignty: national learner data handled under national law; the authority receives de-identified aggregate reports.
  • Fees: the inclusive tier is free for public schools; other terms agreed separately.
  • Nature: framework, non-exclusive; specific projects contracted separately.

4. Use of the samples

Legal notice

The samples in this chapter are for reference by the Association's counsel and local lawyers when drafting and do not constitute legal advice; data-protection, minors, labour and tax clauses must be reviewed by local counsel in each country before signature. All amounts, percentages and terms are completed in the executed text according to negotiation.

10 · Pedagogy & content

Pedagogy and content教研与内容

Course content cannot be generated by technology in place of curriculum work. Curriculum is the effort most likely to become the bottleneck and must start in parallel with engineering. All three pilots use the same content; only the scaffolding language differs.

StageContent goalOutputReview
MVP (end of October)First six weeks of HSK Level 1: pinyin and tones, 150 words, basic sentence patterns, six everyday dialogue scenesSix weeks of courseware and recordings, 300 objective items, 400 review cards, six photo-homework tasks, English scaffoldingChina-side teacher first review → Liu Laoshi final review → entered
Pilot (November–December)All of HSK Level 1 + Level 1 mock examTwelve weeks of content, two mock papers; first drafts of French and Kiswahili scaffoldingAs above, revised with teacher and facilitator feedback
Phase 2 (January–March)HSK Level 2, twelve speaking scenes, French / Kiswahili scaffolding finalisedGraph extended to Level 2; speaking item bankNative-speaker reviewers added
  • Standards alignment: anchored to HSK 3.0 (three bands, nine levels) and YCT; every knowledge point carries level and source in the graph; platform certificates are strictly separate from official certification.
  • Production method: teachers write the syllabus and key explanations; AI drafts exercises and variants; teachers review item by item before entry; unreviewed content never enters the graph.
  • Publishing and Future Academy: modules coded to the Academy's four strands; strong lecture notes and learner work become publishing material; competition prompts serve as writing practice; new titles are listed in the library automatically.
  • Local systems: the pilot collects the three countries' curriculum standards and term rhythms in preparation for certification discussions in Phase 3.
11 · Compliance

Compliance and data合规与数据

The compliance register records law, obligation and evidence per country and becomes a design input only after confirmation by counsel. The table lists boundaries to be confirmed and is not legal advice; applicability in each country follows the PoC-8 legal memos.

ScopeLaw (applicability to be confirmed)Main obligationsSystem implementationStatus
Singapore (Association and core data)PDPAConsent, purpose limitation, protection, retention, transfer, data protection officerPrivacy policy, consent management, Singapore residency, export and deletion endpointsPlatform aligned; the Association appoints a DPO
Nigeria (learners)NDPA 2023 and NDPC rulesLawful basis, guardian consent for minors, transfer conditions, possible registrationGuardian consent, minimisation, transfer records, school sub-tenant isolationLegal memo pending
Kenya (learners)Data Protection Act 2019 (ODPC)Controller / processor registration, protection of minors, transfers with ODPC-recognised safeguardsAs above; registration status in the registerLegal memo pending
Côte d'Ivoire (learners)2013 personal data protection law (ARTCI)Declaration or authorisation of processing, protection of minors, transfer authorisationAs above; filings archivedLegal memo pending
China (remote teachers)PIPL and cross-border transfer rulesMinimisation, notice and consent, transfer assessment thresholdsTeacher data minimised; recording and voice licences; no media node in ChinaLegal memo pending
European community schools (Phase 3)GDPRData processing agreement, data-subject rightsDPA template, export and deletionPhase 3
Content and child protectionNational online child-protection requirementsContent moderation, reporting and appeal, recording noticeReview queue, reporting entry, recording notice announced in three languages at every classBuilt into MVP
AI governanceSingapore AI governance framework (voluntary)Transparency, human oversight, traceabilitySourced outputs, hand-off, audit trail, errors recorded openlyCovered by design principles

Data flows and sovereignty (written into the data schedule)

  • The Association is controller, the technology platform processor, and each school joint controller of its own learners' data; recorded in the MOU and schedule.
  • Core data in Singapore; learner-side local data encrypted; episodic memory append-only, exportable and deletable; deleted within 30 days of account closure.
  • Genuine isolation between schools: tenant-isolation tests are part of the release gate.
  • No commercialisation of data during the pilot; no learner data used to train general models.
  • One local data contact per country (the facilitator may serve) receives data-subject requests in the first instance.
12 · Governance

Governance and communication治理与沟通

Joint programme office

Association lead, platform lead, programme manager, pedagogy lead (Liu Laoshi) and Africa business lead (Ray); 45-minute meeting every Monday; the programme manager keeps the single plan, risk table and contract register.

Decision rights

Scope and timing: programme office; budget and external agreements: Association board; architecture and technology: platform lead; curriculum and teachers: Liu Laoshi; pilot schools and local relations: Ray.

Document system

One shared drive: plan, PRD, ADRs, compliance register, contract register, weekly reports, minutes, acceptance sheets. All external material draws on the brand manual and fact sheet.

CommunicationFrequencyParticipantsFormat
Programme meetingEvery MondayProgramme officeProgress, risks, decisions; 45 minutes
Friday demonstrationEvery FridayAll + school representatives (optional)Real accounts under weak network; 30 minutes
Pilot weekly reportEvery FridayBoard, three schools, teachersOne page: metrics, issues, next week; per school
Board reportOnce at each of M1–M4BoardGate conclusions and decision requests
School communicationEvery two weeksPrincipal, facilitator, parent representativesCall or visit
Change controlAs neededProgramme managerChange note; changes affecting timing, cost or compliance need programme-office approval
13 · Decisions

Decision list决策清单

#DecisionRecommendationDeciderBy
D1Approval and authority to contractBoard resolution authorising the Association lead to sign C1–C7Association boardSeptember, week 1
D2Three pilot countriesNigeria, Kenya, Côte d'Ivoire; alternates Ghana, Tanzania, SenegalAssociation boardSeptember, week 2
D3One pilot school per countryFrom the four-dimension shortlist, recommended by RayAssociation leadSeptember, week 4
D4MVP scope (native app deferred)Per the table in §05Programme officeM1
D5Live / recording emphasis per countryPer PoC-1 conclusionsPlatform leadM1
D6Media or access node in ChinaNone during the pilotPlatform lead + counselM1
D7Unified education brand name“Tongwen” or “Shuqiao”, fixed after trademark search; “HanQiao” remains a code nameAssociation boardEnd of September
D8Pilot feesEntirely free; seats and corporate classes charged from Phase 2Association leadM1
D9Governing law and forumSingapore law + SIAC, enforceability confirmed by local counselAssociation lead + counselOctober, before C2
D10Expansion list and countriesPer cross-country conclusions and the four-dimension scoreAssociation boardM3
14 · Appendix

Appendix · Templates附录 · 模板

A. Principal interview guide (60 minutes)

  1. School profile: grades, enrolment, current Chinese provision, weekly hours and slots.
  2. Devices and network: number and condition of school devices, operator and tariffs, power-cut frequency, shared classrooms and projector, room for a cache-point device.
  3. Teachers and facilitators: who organises the class; share of mother tongue versus English / French / Kiswahili.
  4. Parents: channels, payment habits and capacity, attitude to children's data.
  5. Outcomes: what parents and the school value most (certificate, exam, speaking); awareness of HSK / YCT.
  6. Concerns: about AI, about data, about time and cost.
  7. Cooperation: what the school will provide, what it expects from the Association, expectations of the free pilot and afterwards.

B. PoC record fields

Card number · country and school · date and time slot · location and operator · device · executor · steps · raw measurements · pass / fail · evidence files (screenshots / logs / audio) · conclusions and recommendations · reviewer.

C. Pilot weekly report template (one page, per school)

This week's numbers (active, attendance incl. replay, practice, homework, accuracy, hand-off, cost) · three things that went well · three issues and owners · next week's plan · decisions needed from the board or the school.

D. School scoring sheet

Country · school · contact · network and power (1–5) · demand and organisation (1–5) · partner maturity (1–5) · compliance feasibility (1–5) · total · visit date · conclusion (sign / alternate / exclude).

E. Contract register fields

Instrument ID · type · parties · country · version · signing date · effective date · expiry · data-schedule version · localisation variance table · storage location · renewal reminder.

F. Glossary

PoC
Proof of concept: the smallest experiment that answers one risk question
MVP
Minimum viable product: the smallest feature set that lets a real class complete a full cycle
PWA
Installable web application that runs offline
RTC
Real-time audio and video
TIDAR KAG
The platform's provenance-bound knowledge-graph generation, used for low-hallucination Q&A and mastery judgement
Agent Harness
The base for building, running, evaluating and governing agents
MOU / LOI
Memorandum of understanding / letter of intent
DPA
Data processing agreement or data protection schedule
ADR
Architecture decision record
HSK / YCT / HSKK
Chinese Proficiency Test / Youth Chinese Test / HSK Speaking Test