中文版
1. 我们是谁
本政策说明个人卖家 韩宇翔（Yuxiang Han）（「我们」「运营者」，位于中华人民共和国）在提供橙子超市批发 API时如何处理信息。
当前批发模型为运营者自营 GLM-5.3（H200，FP8），部署在中国大陆。本首发路径不经过火山引擎或其他第三方大模型托管 API。
OpenRouter 等聚合平台是我们的客户（API 调用方）。我们为返回推理结果而处理其发来的提示词。OpenRouter 不是我们的子处理方；终端用户通常首先适用聚合平台自己的隐私政策。
2. 适用范围
- 批发接口 API 请求（如
/or/v1下 OpenAI 兼容路由）； - 我们签发的 API 凭证；
- 运维日志、计量/对账记录与支持沟通；
- 本站相关法律页面。
3. 处理的数据
| 类别 | 示例 | 目的 |
|---|---|---|
| 客户联系信息 | 姓名/称呼、账单邮箱、收款信息 | 履约、结算、支持 |
| API 凭证 | API Key、标识、密钥摘要 | 鉴权、防滥用 |
| 请求元数据 | 时间、模型、token、状态码、请求 ID、IP、限流计数 | 交付、计量、安全、对账 |
| 提示词与生成内容 | 消息、工具调用、模型输出 | 仅用于推理 |
| 支持内容 | 发至联系邮箱的邮件 | 客户支持 |
4. 提示词、日志与训练（重点）
- 推理：提示词与生成内容仅为完成推理并返回结果。
- 训练：我们不会使用此类内容训练、微调或改进基础模型（含 GLM-5.3），也不会为训练目的出售。
- 正文日志：默认尽量不在请求结束后长期保存完整正文；出于安全/滥用/故障调试，内容日志可能保留最长7 天，其后删除或匿名化（进行中事件、争议或法律要求除外）。
- 元数据：为计量与运维保留更久（见第 6 节）。
5. 处理目的
履约批发推理、计量对账、安全保障，以及遵守适用中国法律法规并响应合法要求。聚合平台应自行确保向中国大陆上游发送内容具备合法基础。
6. 数据保留
| 数据 | 期限 |
|---|---|
| 提示词/生成正文（如有记录） | 最长 7 天（例外见上文） |
| 推理过程短暂缓冲 | 仅单次请求期间 |
| 请求元数据与 token 计量 | 最长 24 个月；用量记录满 24 个月自动删除。容器运行日志（不含提示词与输出）由运维轮转清理，目标上限 24 个月，按季度人工核查 |
| API Key 与客户联系信息 | 合作期间 + 结束后最长 24 个月 |
| 安全/滥用调查记录 | 最长 24 个月（调查未结束可延长） |
| 支持邮件 | 线程末封后最长 24 个月 |
7. 子处理方与基础设施
模型推理：运营者在中国大陆自有 H200（FP8）。本首发模型不使用火山引擎等第三方托管 LLM API。
基础设施：计算/网络/存储由中国大陆云计算基础设施提供商提供；未授权其将提示词用于训练。
draft-0.1 未再引入可接触提示词正文的其他内容型子处理方。OpenRouter 等聚合平台就其提交的请求属于独立客户，不列为我们的子处理方。
8. 数据驻留
主要处理与存储位于中国大陆。境外聚合平台向本 API 发送内容由其发起；本服务暂无境外独立推理区域。
9. 安全
采用与个人运营服务相称的措施：HTTPS/TLS、API Key、主机访问控制、限流，及在已实现范围内隔离批发凭证。无法保证绝对安全。
10. 您的请求
批发客户可申请轮换/吊销 Key、导出其计量元数据、在法定与结算允许范围内删除联系信息与密钥，或询问是否启用正文日志。联系：hyx_8100@hotmail.com。
聚合平台终端用户请优先联系该平台。
11. 未成年人
本服务面向企业与开发者。请勿在缺乏合法基础时故意处理儿童个人信息。
12. 变更
我们可能更新本政策并变更文首日期/版本。涉及日志或训练立场的重大变更将在本页披露。
13. 联系
运营者：韩宇翔（Yuxiang Han）（个人）
邮箱：hyx_8100@hotmail.com
English
1. Who we are
This policy describes how 韩宇翔（Yuxiang Han） (“we”, “Operator”), an individual seller in the PRC, processes information when providing the Orange Supermarket Wholesale API.
The Service currently offers operator-run GLM-5.3 inference (H200, FP8) in mainland China. This first wholesale path does not use Volcano Engine or other third-party hosted LLM APIs.
Aggregators such as OpenRouter are our customers (API consumers). We process prompts they send in order to return completions. OpenRouter is not our subprocessor for those prompts. End users typically rely first on the aggregator’s privacy notice.
2. Scope
- Wholesale API requests (OpenAI-compatible routes under
/or/v1); - API credentials we issue;
- Operational logs, metering/billing records, and support;
- Related legal pages on this host.
3. Data we process
| Category | Examples | Purpose |
|---|---|---|
| Customer contact | Name/handle, billing email, payout details | Contract, billing, support |
| API credentials | Keys, key IDs, hashed secrets | Auth, abuse prevention |
| Request metadata | Time, model, tokens, status, request id, IP, rate limits | Delivery, metering, security, invoicing |
| Prompt & completion content | Messages, tool payloads, outputs | Inference only |
| Support content | Emails to our contact address | Support |
4. Prompts, logging, training
- Inference: Content is processed solely to provide the requested inference.
- Training: We do not use prompts/completions to train, fine-tune, or improve foundation models (including GLM-5.3), and we do not sell such content for training.
- Content logs: We aim not to retain full bodies after the request; for security/abuse/debug, content logs may be kept up to 7 days, then deleted or anonymized (except active incidents, disputes, or legal requirements).
- Metadata: Retained longer for billing and operations (see §6).
5. Purposes
Perform the wholesale contract; meter and reconcile usage; secure the Service; comply with applicable PRC law. Aggregators should ensure a lawful basis before sending end-user content to a mainland China upstream.
6. Retention
| Data | Period |
|---|---|
| Prompt/completion content (if logged) | Up to 7 days (exceptions above) |
| Transient inference buffers | Request lifetime only |
| Request metadata & token metering | Up to 24 months; usage records are deleted automatically after 24 months. Container operational logs (no prompts or outputs) are rotated by operations with a target maximum of 24 months, reviewed quarterly |
| API keys & customer contact | Relationship + up to 24 months after |
| Security/abuse records | Up to 24 months (longer if open) |
| Support email | Up to 24 months after last message |
7. Subprocessors & infrastructure
Model inference: Run by the Operator on H200 (FP8) in mainland China. No Volcano Engine / third-party hosted LLM API for this first model.
Infrastructure: Compute/network/storage via cloud infrastructure providers in mainland China; not authorized to use prompts for training.
No other content subprocessors for GLM-5.3 wholesale at draft-0.1. Aggregators such as OpenRouter are independent customers, not our subprocessors.
8. Data residency
Primary processing and storage are in mainland China. Cross-border submission by foreign aggregators is initiated by them; we do not operate a separate non-China inference region today.
9. Security
Reasonable measures for an individually operated service: HTTPS/TLS, API keys, host access control, rate limits, and wholesale credential separation where implemented. No system is 100% secure.
10. Requests
Wholesale customers may rotate/revoke keys, request export of their metering metadata, request deletion of contact/keys subject to legal/billing retention, or ask whether content logging is enabled. Contact: hyx_8100@hotmail.com.
Aggregator end users should contact the aggregator first.
11. Children
The Service is for businesses and developers, not children.
12. Changes
We may update this policy and the effective date/version. Material changes to logging or training posture will be reflected here.
13. Contact
Operator: 韩宇翔（Yuxiang Han） (individual)
Email: hyx_8100@hotmail.com