Latent is attentioned.

Invert the visual hierarchy把视觉层级倒过来

Nearly every product in the field — including the systems hailed as the frontier of Generative UI — is fundamentally tool-call-rooted: the UI is generated from tool output, and reasoning is squeezed into a footnote beside the action, an after-the-fact justification. Latent inverts this hierarchy. Actions recede into a peripheral, recallable trace; the agent's evolving understanding takes the main stage.

领域里几乎所有产品——包括被视为最前沿的 Generative UI——本质上都是 tool-call-rooted：UI 的生成源头是工具的输出，推理被压成动作旁边的一行小字、一个事后的辩护。潛 把这个层级倒过来。动作退为外围、可调取的痕迹；agent 不断演化的理解占据主舞台。

Three zones, one proportion三个区域，一种比例

Every agentic interface decomposes into three attention zones. Their visual weight is a hard constraint of the design language, not a layout preference.

每个 agentic 界面被分解为三个注意力区域。它们的视觉权重是设计语言的硬约束，不是布局偏好。

The spine against reasoning theater对抗 reasoning theater 的脊梁

This is what sets Latent apart from every existing approach, and it is its safety floor. The rendered latent is not the real latent — chain-of-thought is often after-the-fact rationalization, and the prettier it is, the more easily it gilds unfaithful cognition with the gold leaf of "trustworthy." So: only falsifiable, evidence-grounded cognition deserves the main attention. Anything that wants the main stage must be one of three legitimate forms.

这是潛 区别于一切现有方案的核心，也是它的安全底座。渲染出来的 latent 不等于真实的 latent——链式推理常是事后合理化，做得越漂亮越容易给不忠实的认知贴上“可信”的金箔。因此：只有可证伪、被证据锚定的认知，才配占据主注意力。任何想进入主舞台的元素，必须是以下三种合法形态之一。

This contract is exactly your five-layer diagnostic ontology: the upper ontological layers (hypotheses, conclusions) are the subject on stage, but every one of them must have a thread tied back down to the Observable Primitives below. Reasoning theater is dangerous precisely because its narrative floats free; the Grounding Contract uses "must be falsifiable, must land on evidence" to keep free-floating assertions off the main stage.

这条契约正是你的五层诊断本体论：本体论上层（假设、结论）是台前主体，但每一个都必须有一根线栓回下层的 Observable Primitives。Reasoning theater 之所以危险，正因为它的叙述悬空；Grounding Contract 用“必须可证伪、必须落到证据”把悬空的断言挡在主舞台之外。

Color is epistemic state颜色即认识论状态

Latent does not use color for decoration. Color is a meaning-bearing dimension: it encodes the epistemic state of a cognitive element. See gold and you know "grounded / fairly certain"; see blue and you know "the model changed its mind" — blue is a cue color (worth noticing), not a red error. Changing your mind is a good thing and shouldn't be treated as an error warning.

潛 不用颜色装饰。色彩是承载意义的维度：它编码一个认知元素的认识论状态。看到金色就知道“已锚定/比较确定”，看到蓝色就知道“模型改了主意”——蓝是提示色（值得注意），不是红色的报错。改主意是好事，不该被当成错误警告。

Three voices三种声音

Type is divided by role: the agent's cognitive "voice," the interface's functional text, and evidence / provenance as instrument readouts. The voice uses serif, because thought should look deliberated; evidence uses monospace, because it should look like a verifiable instrument readout.

排版按角色分工：agent 的认知“声音”、界面的功能文本、以及作为仪器读数的证据/溯源。声音用衬线，因为思考应当显得被斟酌过；证据用等宽，因为它应当显得像可核验的仪器读数。

A language of depth: surfacing and sinking深度的语言：浮现与沉降

Motion serves the metaphor: latent is depth (潛). Understanding surfaces from the depths; confidence settles; refuted hypotheses sink. Movement is never mere decoration — it narrates the transition of epistemic state. (hover to see)

动效服务于隐喻：latent 即深度（潛）。理解从深处浮现；置信度凝定；被推翻的假设沉降。运动从不只是装饰——它讲述认识论状态的迁移。（悬停查看）

Core components核心组件

The components below are rendered live by @latent/react; every prop is a z.infer<> subtype of @latent/schema. Open any provenance to see "de-emphasized ≠ hidden" — the tool call recedes visually, but is always one click away and auditable.

下列组件由 @latent/react 实时渲染，每个 prop 都是 @latent/schema 的 z.infer<> 子类型。点开任意 provenance 看“弱化 ≠ 隐藏”——工具调用在视觉上退后，但永远一键可调、可审计。

What the language looks like in motion语言运转起来的样子

A true-to-form network incident diagnosis, driven by a validated CognitiveState instance. The user story above sets out who, why, and what's being done here; by default the interface shows the current understanding the SRE sees on arrival (conclusion first). To watch how it formed, hit "Replay" (demo); toggle Trace at the top right to see the "two personas."

一次真实形态的网络故障诊断，由一份校验过的 CognitiveState 实例驱动。上方用户故事交代谁、为什么、在这里做什么；界面默认展示 SRE 到达时看到的当前理解（结论先行）。想看它怎么形成，点「回放」（演示）；右上角切 Trace 看“两种人格”。

Understanding scales with complexity理解态随复杂度伸缩

"For all AI apps" means some tasks are cognitively shallow. The mass of the understanding should scale with the epistemic complexity of the problem — don't put up a hypothesis board for "format this JSON."

“给所有 AI 应用”意味着有些任务认知很浅。理解态的体量应随问题的认识论复杂度伸缩——别给“格式化这段 JSON”摆一块假设板。

Live: one and the same @latent/react + ProportionalView, scaling automatically across three tiers of complexity — low latent collapses to one grounded line + provenance, high latent unfolds the full hypothesis board.

实时：同一套 @latent/react + ProportionalView，在三档复杂度下自动伸缩——低 latent 退为一行 grounded + provenance，高 latent 展开完整假设板。

The 6 archetype demos below spread out exactly along this axis, across different agent types.

下方 6 个原型 demo 正是沿这条轴、跨不同 agent 类型铺开。

One language, across every kind of agent一套语言，跨各种 agent

Diagnosis is just one archetype. The beauty of the Grounding Contract is that it holds across archetypes: a decision, a requirement understanding, an action taken on your behalf that occupies the main stage must equally be grounded (hung on evidence / criteria / your own words), or marked tentative (with "what would change it"), or left as an honest open. Content role is expressed on an orthogonal kind axis (decision / plan / requirement / option / answer / risk…); color still encodes epistemic state.

诊断只是一种原型。Grounding Contract 的妙处在于它跨原型成立：一个决策、一个需求理解、一个替你执行的动作占据主舞台，同样必须 grounded（挂在证据/准则/你的原话上）、或标为 tentative（带“什么会改变它”）、或留作诚实的 open。内容角色由正交的 kind 轴表达（decision / plan / requirement / option / answer / risk…），颜色仍编码认识论状态。

From components, up to a whole app从组件，升到整个 App

The three-zone model scales up to the app level: every agentic product is composed of five attention zones, where visual weight is a hard rule, not a layout preference. The macro inversion — where others put the tool-call stream on the main stage, Latent puts "understanding" there, with artifacts (code / canvas) and activity (terminal / render queue) receding to the periphery.

三区模型放大到 App 级：每个 agentic 产品都由五个注意力区域组成，视觉权重是硬规则，不是布局偏好。宏观的倒置——别家把工具调用流放在主舞台，潛 把「理解」放上去，工件（代码/画布）与活动（终端/渲染队列）退到周围。

Attention weight: Stage > Artifact > { Context, Activity }; Intervention is always reachable. The two real products below are laid out with the same grammar:

注意力权重：Stage > Artifact > { Context, Activity }；Intervention 永远可达。下面两个真实产品用同一套语法布局：

Seven principles七条原则

Epistemic honesty, not theater认识论诚实，而非剧场

The design language also governs how the agent speaks. Latent's voice refuses the "hard at work" performance, and refuses to dress after-the-fact narration up as insight. It states confidence, marks the unknown, and stays honest about "the gap between narration and computation."

设计语言也约束 agent 怎么说话。潛 的语态拒绝“正在努力工作”的表演，也拒绝把事后叙述伪装成洞见。它陈述置信、标注未知、对“叙述与计算之间的落差”保持诚实。

Red lines for adoption采纳红线

When any team adopts Latent, the following are anti-patterns that quietly betray the core of the language. They look like "showing the reasoning," but in fact they rebuild the theater.

任何团队接入潛 时，以下是会悄悄背叛该语言核心的反模式。它们看起来像“在展示推理”，实则在重建剧场。