不可变版本 / Immutable version 2026-10-25.2. 本副本不加载分析或应用脚本 / No analytics or app scripts.
把一次请求的处理、聊天与文件保存、工作区经验、模型训练和公开发布分别说明。关闭一种功能不等于关闭所有处理。
2026-10-25.2 · 2026-10-25
为生成回答、执行授权工具、搜索或媒体任务，必要输入、上下文和结果可经过 Laintas 与所选来源及其必要下游。选择某个模型不是对无限用途的授权。您应有权提交上下文；企业、客户保密或特定驻留资料应先确认部署和接收方。自带 key 仍可能经过网关，费用由您的供应商账户承担；自身收费与数据处理分别判断。
浏览器数据库、桌面文件、远程机器、Laintas 云存储、第三方仓库和 Scene 世界记录分开存在。同步是一份额外副本，退出登录不代表删除。只有某一模块不保存正文，也不能推定整个调用链零留存；例如 Issues 私密流返回到主站对话后，聊天记录仍可按账户设置保存。
网关经验用于同一账户工作区的跨会话延续，与训练共享开关无关。当前实现可从 CLI、CLI Windows 与 Helpwo 已完成轮次提取用户表达、行动摘要和结果，经摘要后在后续上下文中使用；并非只存用量元数据。它可能涉及额外摘要模型处理，不承诺原始秘密绝对不会进入摘要。可通过产品已提供的管理功能或联系客服请求查询、更正、删除及停止不必要处理；本地记忆与网关经验需分别处理。
当前云端训练机制要求服务端记录有效主动选择，只覆盖 CLI 与 CLI Windows 实际经过网关的对应输入和输出，可能含文件片段、工具目录、上下文和任务轨迹关联信息。/training status 查看、/training on 开启、/training off 停止新采集；/training local 是另一个本地采集开关。请在开启前理解范围并确认所有相关资料可授权，不将雇主或客户禁止提供的内容纳入。身份脱敏和密钥清洗降低风险，不保证所有文本已完全匿名化。
公开问题和被采纳回答具有与 CLI 不同的授权路径：适用 Issues 专用条款中发布或答题前明确提供并有效接受的许可，在合法范围筛选和用于训练，不自动受 CLI 训练开关控制。仅公开署名为匿名不等于内容匿名化。密封人工任务上下文、Issues 私密流和无权授权的第三方资料不因该许可成为训练材料。不能追溯把新许可套用于未获得相应有效授权的旧内容。涉及个人信息时另按法律确定依据和所需同意。
当前网关自带 key 路径不进入 Laintas 的经验和训练采集；Compute 自有机器转发不持久保存推理正文在其调用明细表。组织和自建后端按具体路径处理，不概括继承个人 CLI 训练选择。但发起请求的客户端可以另存会话，您选定的供应商或设备也可有独立记录。这些旁路不免除必要用量、费用、安全与法律记录。
停止新采集、删除原始记录、删除已导入训练集和撤回已训练模型影响是不同事项。关闭开关不自动删除既有数据，但不排除适用法律赋予的删除、限制等权利。可提供账户、来源、时间范围和需要处理的内容联系客服，勿为证明而发送更多不必要的秘密。对可识别样本和衍生副本依法核查；已训练模型不能保证逐条撤回影响，不把技术困难当作当然免除法律义务。
推理过程中的临时数据、用量账本、工作区经验和训练集有不同期限。训练原始台账的清理只针对符合条件的记录，导入后数据集另按项目用途和法律要求保存，不能把原始台账的周期当作全部训练数据上限。保存应限于有效用途的必要时间，安全清洗或人工审核可能接触内容，并受权限和用途限制。您可按隐私政策请求具体范围和依据；对新增用途依法另行处理告知或同意。
下表帮助区分主要路径，具体范围仍以本说明及相应产品规则为准。关闭训练不自动关闭聊天同步、工作区经验、公开发布或第三方处理；反过来，购买会员和保存会话也不等于主动加入训练。
| 用途 | 内容与触发 | 控制和边界 |
|---|---|---|
| 完成请求 | 当前请求、必要上下文与工具结果 | 按功能与来源处理；不是训练授权 |
| 保存与同步 | 本机或所选云存储的历史与文件 | 分位置删除；不等于全链路零留存 |
| 网关经验 | 支持路径的已完成轮次与摘要 | 用于账户延续；与训练开关分开 |
| CLI 训练 | 服务端有效主动选择后的对应调用 | 关闭停止新采集；旧样本另处理 |
| Issues 公开训练 | 具有有效许可的公开问题与被采纳回答 | 独立发布规则；不含密封工作 |
| 安全与争议 | 处理事件所需的最少记录 | 限制目的、权限及依法保留范围 |
提交差评、报错、退款材料或侵权投诉，允许为解决该事项处理必要内容，不自动把完整会话变成通用模型训练材料。若某个反馈入口另请求用于训练，应明确范围并依法取得所需授权。防欺诈、内容安全和法定留存可以有独立处理依据，但不得把所有普通对话无限扩张解释为安全样本。进行去标识化仍可能保留可识别文本；真正匿名化须达到依法不能识别且不能复原的要求，不能仅因去掉用户名就这样称呼。
Separate explanations for request processing, saved chats and files, workspace experience, model training and publication. Disabling one does not disable every kind of processing.
2026-10-25.2 · 2026-10-25
Necessary inputs, context and results may pass through Laintas, the chosen source and necessary downstream to answer, use authorised tools, search or process media. Choosing a model does not authorise unlimited purposes. Ensure authority over context and confirm deployments and recipients for confidential or residency-bound material. Own keys can still use the gateway and charge your provider account; charging and processing are separate questions.
Browser databases, desktop files, remote machines, Laintas cloud storage, third-party repositories and Scene worlds are separate stores. Sync creates another copy and sign-out does not delete it. Non-retention in one component does not mean a zero-retention chain: an Issues private answer returned to main-site chat can still be saved as chat history.
Gateway experience supports continuity within an account workspace, separately from training sharing. Current code can derive user statements, action summaries and outcomes from completed CLI, CLI Windows and Helpwo turns, summarise them and reuse them in later context. It is not merely usage metadata and may use additional summarisation models; summaries are not guaranteed to exclude every secret. Use available controls or support for access, correction, deletion and stopping unnecessary processing. Manage local memories and gateway experience separately.
Current cloud training requires a current server-side opt-in and covers the corresponding CLI and CLI Windows inputs and outputs actually observed by the gateway, potentially including excerpts, tool catalogues, context and trajectory links. Use /training status, /training on and /training off; /training local is separate local capture. Understand the scope and ensure authority before opting in, including employer and client restrictions. Pseudonymisation and credential scrubbing reduce risk but do not guarantee complete anonymisation.
Public questions and accepted answers follow a different authority from CLI: the Issues licence expressly provided and validly accepted before posting or answering permits lawful selection and training and is not automatically controlled by the CLI switch. Anonymous display names do not anonymise content. Sealed task context, Issues private streams and unlicensable third-party material are excluded. New permissions do not apply retroactively to older content lacking valid authority. Personal information needs its own lawful basis and consent where required.
Current own-key gateway paths bypass Laintas experience and training capture. Own-machine Compute call-detail tables do not persist inference bodies. Organisation and self-hosted paths follow their deployment and do not generally inherit a personal CLI training choice. The requesting client, chosen provider or device can still keep its own history. These paths retain necessary usage, billing, security and legal records.
Stopping future capture, deleting raw records or imported samples and reversing a trained model’s influence are different. Switching off does not automatically delete earlier data, without excluding legal deletion or restriction rights. Contact support with account, source, time range and scope, without sending unnecessary secrets as proof. Identifiable samples and derived copies are reviewed under law. Per-sample reversal of trained effects cannot be guaranteed, and technical difficulty does not automatically remove legal duties.
Transient inference, usage ledgers, experience and datasets have different retention. Raw-ledger pruning only affects eligible rows; imported datasets follow project purpose and law, so the ledger interval is not a universal training-data limit. Keep data only as needed for valid purposes. Security scrubbing and authorised review may expose content, within restricted access and purposes. Request scope and basis under the Privacy Policy; new purposes require applicable notice or consent.
This table distinguishes major paths; this notice and product rules define their scope. Disabling training does not automatically disable chat sync, experience, publication or third-party processing. Buying membership or saving a conversation does not itself opt you into training.
| Purpose | Content and trigger | Controls and limits |
|---|---|---|
| Fulfil a request | Current request, necessary context and tool results | Feature/source processing; not training authority |
| Storage and sync | History/files in local or chosen cloud storage | Delete per location; no end-to-end zero-retention implication |
| Gateway experience | Completed turns and summaries on supported paths | Account continuity, separate from training switch |
| CLI training | Matching calls after valid server-side opt-in | Off stops new collection; prior samples need separate handling |
| Issues public-content training | Public questions/accepted answers with valid permission | Separate publication rules; sealed work excluded |
| Security and disputes | Minimum records needed for a case | Restricted purpose, access and lawful retention |
Negative ratings, bug reports, refund evidence and rights complaints permit necessary handling of that matter, not automatic general training on the whole conversation. A feedback flow requesting training must identify scope and obtain required authority. Fraud prevention, safety and legal retention may have distinct grounds but must not reclassify every ordinary conversation as an unlimited safety sample. De-identification can leave identifiable text; anonymisation requires the applicable standard of non-identifiability and irreversibility, not merely removing a username.