Local-first direction本地优先
Personal context and everyday intelligence are intended to remain on user-controlled devices when practical.
在功能允许的情况下，个人上下文与日常智能尽可能运行在用户可控制的设备上。
L.I.A. · PERSONAL INTELLIGENCE SYSTEM
Intelligence that stays close.一个始终在身边的智能系统。
L.I.A. is CAELARIUM's local-first personal-intelligence foundation. It is designed to keep one consistent relationship while conversation, memory, Skills, tools and specialist systems evolve underneath.
L.I.A. 是 CAELARIUM 的本地优先个人智能基础。它希望在底层对话、记忆、Skill、工具与专家系统不断升级时，仍然保持同一个稳定入口与关系。
AWAKENED · ACTIVE DEVELOPMENT
WHAT L.I.A. ISL.I.A. 是什么
One relationship. Many capabilities underneath.一个稳定关系，底层承载多种能力。
L.I.A. is not defined as one giant model. The system separates identity and interaction from modular capabilities that can be tested, permissioned, replaced and improved independently.
L.I.A. 并不被定义成一个万能大模型。人格与交互关系和底层模块化能力分离，每项能力都可以独立测试、授权、替换和升级。
Modular Skills模块化 Skill
Document, table, research, coding and future system abilities can be separated into explicit Skills with their own inputs and checks.
文档、表格、研究、代码与未来系统能力可以拆分为输入和验证明确的独立 Skill。
Permission before action行动之前先授权
The direction is to constrain intent, check permissions, act, observe the real result, verify it and only then report completion.
目标会先被约束和拆解，执行前检查权限，执行后重新观察现实结果，再验证并汇报。
WHAT EXISTS TODAY今天已经存在的部分
The first layer is presence.第一层，是它真的在那里。
The current L.I.A. prototype can wake by name, listen through local speech recognition, answer with a consistent local voice, keep a conversation alive, move between cold and warm sleep, and run its fast conversational brain on the user's own machine.
当前 L.I.A. 原型已经能够本地叫名唤醒、本地语音识别、以稳定本地声线回答、维持连续对话，并在冷休眠与热休眠之间切换；负责日常交流的快脑也运行在用户自己的电脑上。
Capability boundary能力边界
The broader architecture below is a mix of implemented foundations and product direction. A layer is not presented as shipped merely because it appears in the architecture.
下面更完整的架构同时包含已经实现的基础与正在构建的方向。出现在架构里，并不等于已经作为完整产品能力发布。
TEN LAYERS十个层次
A path from conversation to trustworthy action.从对话，走向可以被信任的行动。
Local Core
Wake, listen, respond and keep everyday interaction available locally.
本地唤醒、听懂、回答，并让日常交互能够持续存在。
Quality Gate
Separate “the chain ran” from “the answer was correct” through repeatable evaluation and correction.
把“流程跑通”和“答案正确”分开，通过可重复评测与纠错来判断可信度。
Skills
Keep capabilities modular so each one can be permissioned, tested and improved independently.
把能力拆成模块，让每项能力都可以独立授权、测试和升级。
World State
Represent time, device state, apps, files, projects, tasks, permissions and current activity as structured state.
把时间、设备、应用、文件、项目、任务、权限与当前活动变成结构化状态。
Long-term Memory
Keep durable memory outside model context and design it to remain viewable, correctable, deletable, exportable or disabled.
长期记忆独立于模型上下文，并以可查看、修改、删除、导出或关闭为设计方向。
Action + Verification
Act only inside permission boundaries, then observe and verify the real result before reporting success.
只在权限边界内行动，并在汇报成功前重新观察与验证真实结果。
Router + Planner
Convert natural-language goals into constrained intent, steps, dependencies and the right Skill or agent.
把自然语言目标转成受约束意图、步骤、依赖，并交给正确的 Skill 或智能体。
Specialists
Let document, coding, research, computer and other specialists work behind one consistent L.I.A. relationship.
让文档、代码、研究、电脑等专家能力在同一个 L.I.A. 关系背后协作。
Identity + Learning
Keep identity, values, voice and boundaries separate from any single foundation model so the brain can evolve without resetting the relationship.
把身份、价值、声线和边界与单一基础模型分离，让大脑升级而不重置关系。
Derivative Worlds
Reuse proven foundations in specialized worlds such as L.I.A. Recall while keeping explicit permissions and honest capability boundaries.
让 L.I.A. Recall 等专业世界复用成熟基础，同时继续保持明确权限与真实能力边界。
A DERIVATIVE WORLD一个衍生智能世界
L.I.A. Recall applies the foundation to one commercial problem.L.I.A. Recall 把这套基础能力应用到一个明确的商业问题。
Recall focuses on recovering value from existing business opportunities while inheriting the same direction around permissions, bounded Skills, controlled memory and verifiable outcomes.
Recall 专注从企业已有商业机会中追回被忽略的价值，同时继承权限明确、Skill 有边界、记忆可控和结果可验证的方向。