@chrono-meta/fh-gate 2.7.0 → 2.8.0

package/README.md

@@ -18,16 +18,21 @@
 </p>
 
 <p align="center">
-  <b>Ask it for things. When the asking repeats, it offers to build you the thing.</b>
+  <b>A meta-harness with the quality gates built in.</b>
+</p>
+
+<p align="center">
+  Projects, skills, harnesses — building them, checking them, speeding them up: you ask for it here.<br>
+  It does not simply hand the result back. It puts the work past several checks that fail in
+  <i>different</i> ways first.<br>
+  <b>And when the same request keeps coming back, it builds you the harness that does it for you.</b>
 </p>
 
 <p align="center">
   You already tell Claude Code the same things over and over — the checks to run, the rules to hold,
-  the shape a change has to have.<br>
-  <b>forge-harness turns that into something reusable</b>: skills, gates and agents that live in your
-  repo and fire on their own.<br>
-  Its skills stay general on purpose and get shaped to your case as you go.
-  <b>When one shape keeps coming back, it offers to ship it</b> as its own skill, or its own harness.
+  the shape a change has to have. That is the part that becomes reusable, and it keeps its general
+  form on purpose so it can be shaped to your case as you go.<br>
+  <sub>What grows is the number of attempts: the trial and error comes off you and runs in parallel.</sub>
 </p>
 
 ---
@@ -41,24 +46,40 @@ git clone https://github.com/chrono-meta/forge-harness.git ~/projects/forge-harn
 cd ~/projects/forge-harness && claude
 ```
 
-**Then type `hi`.** A numbered menu appears and takes it from there — pick a door, answer a couple of
-questions, and it runs the install wizard for you. Everything below this line is reference for when you
-want it, not homework before you start.
+**Then type `hi`** — or a greeting in whatever language you actually think in: `안녕`, `こんにちは`,
+`你好`, `hola`, `bonjour`. **Any of them opens the menu**, and it will try to answer in the language you
+used. A numbered menu appears and takes it from there — pick a door, answer a couple of questions, and
+it runs the install wizard for you.
+
+<sub>🟥 <b>Honest about that last part</b>: matching your language is a prose rule with nothing
+mechanical behind it, so it does not always hold. Measured blind at the floor tier on 2026-08-21, on a
+<b>clean clone like the one you just made</b>: Chinese and Korean greetings both came back fully
+translated, door labels included. On the maintainer's own machine — which pins a default language —
+Chinese landed only 1 time in 5, which is why this note exists at all. What is still shaky is whether
+the menu fires: one greeting variant produced no menu. If it answers in the wrong language, or skips
+the menu, just say so — it will switch. Written up in <code>CLAUDE.md</code> §Voice/Tone rather than
+smoothed over.</sub>
+
+Everything below this line is reference for when you want it, not homework before you start.
+
+- **What it amplifies** — the number of attempts. Trial and error moves off you and runs in parallel.
 
-**What it amplifies** — the number of attempts; trial and error moves off you and runs in parallel.
-**What it does not** — the model's ceiling. A harness lifts a model to its own ceiling, not past it.
-**How you can check** — it grades itself in public, five identities, in every
-[release](https://github.com/chrono-meta/forge-harness/releases); the ones that are not green name
-the real run still missing.
+- **What it does not** — the model's ceiling. A harness lifts a model to its own ceiling, not past it.
+
+- **How you can check** — it grades itself in public. It names **five things it claims to be**
+  (harness cluster · project incubator · governance gate · frontier → org propagation · amplifier) and grades each
+  one honestly in every [release](https://github.com/chrono-meta/forge-harness/releases). Any square
+  that is not green names the real run that is still missing.
+  <sub>Each of the five is spelled out further down, under «The five identities».</sub>
 
 ---
 
 <p align="center">
-  <img src="docs/pillars.svg" alt="FORK - ADAPT - COLLABORATE - EMPOWER" width="680">
+  <img src="docs/pillars.svg" alt="HARNESS - FORGE - ACCELERATE - COMPOUND" width="680">
 </p>
 
 <p align="center">
-  <b>Quality is the lever; speed is the result.</b> <i>Fork it. Rename it. Make it yours.</i><br>
+  <b>Quality is the lever; speed is the result.</b><br>
   <sub>If this is useful, a star helps others find it.</sub>
 </p>
 
@@ -118,7 +139,9 @@ cd ~/projects/forge-harness
 claude
 ```
 
-> ✅ Then **type a greeting ("hi")** — the 🐿️ door menu appears on a typed greeting, not on launch alone.
+> ✅ Then **type a greeting** — `hi`, `안녕`, `こんにちは`, `你好`, whatever you actually think in; it
+> tries to reply in that language (see the note above — it is right more often than not, not always).
+> The 🐿️ door menu appears on a *typed* greeting, not on launch alone.
 > Say **"Connect a project"** → hub scans `../`, finds `.git` directories, creates `tracks/{project}/`.
 > For full initial setup (hooks · gates · baseline — each item individually approved, declining is
 > respected and recorded), ask for **`/install-wizard`**.
@@ -127,8 +150,8 @@ claude
 
 **Your first 15 minutes** — what success looks like, and what to do with it:
 
-1. You'll know setup worked when a greeting ("hi") shows the 🐿️ door menu, and "Connect a project"
-   creates `tracks/{your-project}/`.
+1. You'll know setup worked when a greeting — in any language — shows the 🐿️ door menu, and
+   "Connect a project" creates `tracks/{your-project}/`.
 2. Then grab an immediate win in the same session: say **"accelerate this project"** (ranked plan of
    skills/plugins worth wiring, install-gated) or **"run /context-doctor"** (token-waste scan).
 3. One honest note: FH's core payoff is **compounding** — session records, harvested learnings,
@@ -347,11 +370,25 @@ harness is actually used.
 
 **The three-stage process** — this is an *order of investment*, not a menu:
 
+```mermaid
+flowchart LR
+  S["① Circuit<br/>before design"]
+  P["② Parallel decorrelation<br/>in the middle"]
+  B["③ Burn it down<br/>on six axes"]
+  A(["⟹ It accelerates"])
+  S --> P --> B --> A
+  style A fill:#0f766e,stroke:#0f766e,color:#fff
+```
+
+**Speed is the arrow at the end, not a fourth box.** And ② is **two dials, turned separately** —
+*decorrelation* (reacts to blind-spot risk: a different model family ⓐ, a different standpoint ⓑ)
+and *parallelism* (reacts to surface size: can one context hold it?). Do not multiply them, choose.
+
 ```
 ① Circuit before design   the judgment circuit goes in FIRST — success · leaning · out-of-scope ·
                           never-do — not written up afterwards as a record of what you did
 
-② Decorrelate in the      split the work into checks that fail DIFFERENTLY and run them at once.
+② Parallel decorrelation  split the work into checks that fail DIFFERENTLY and run them at once.
    middle, to accelerate  Choose which differences matter — a second reviewer of the same kind is
                           not decorrelation, it is the same blind spot twice. Parallelism has no
                           direction of its own; the judgment circuit from ① is what picks.
@@ -362,6 +399,12 @@ harness is actually used.
                           riding it does not make that axis see what it cannot see
 ```
 
+> 📖 **From here down is for whoever wants to go further — it is not needed to start using this.**
+> If you came to install it and get going, the two-minute section at the top is the whole job; you can
+> stop here and come back when a check surprises you. What follows is the reasoning behind the gates,
+> written for someone already running the hub. Unfamiliar words →
+> [`GLOSSARY.md`](knowledge/shared/GLOSSARY.md).
+
 **The six verification axes** — where "we reviewed it" usually turns out to mean only the first of them.
 
 🟥 **Axes are not divided by *how adversarial* they are. They are divided by *what they were given*.**
@@ -392,6 +435,22 @@ That is why the column that matters most below is *what it gets*:
 > beforehand vs claimed afterwards) is a **posture**, like adversariality, not an axis. Minting a
 > new one would have repeated the exact error the blind reclassification above found.
 
+> **A peer session counts on ⓑ and ⓓ — decided 2026-08-21, and *which* peer you ask is the whole
+> trick.** A parallel session of this same harness, running hot on a different branch of the work, is
+> not a copy of you. At the point it got hot it has genuinely grown a second face: a real standpoint
+> (ⓑ) and a real someone-else's-codebase (ⓓ). So peer findings are recorded on those two axes — no
+> seventh axis was minted, and the marker's `axes-run` alphabet did not change.
+>
+> This is why the ⓓ row above says *someone else's codebase* rather than *another company's repo*:
+> the boundary that matters is **whose working context produced the judgment**, not whose GitHub org
+> owns the files. A peer hot on a different branch is across that boundary; a subagent you spawned
+> from this context is not, however different the repo it reads.
+>
+> 🟥 The corollary is the part that bites. **Ask the peer about the axis it actually got hot on.**
+> Anywhere else it wears your face and decorrelates nothing — same input, same blind spot. And a
+> subagent cannot stand in for it, nor can re-reading your own work: the second face comes from that
+> session having *done* different work, and a prompt cannot manufacture it.
+
 **You do not run all six every time, and that is the design** — do not multiply them, **choose**:
 
 ```
@@ -414,8 +473,9 @@ itself.
 **Why this is not superseded by base-model advances** — an axis is defined by its **input**, not by the
 *reviewer's ability*. A stronger model still **cannot see information it was not given.** Scaffolding
 sheds as models improve, but **input-boundary decorrelation does not**, and a single author cannot, by
-definition, step outside their own input. 🟥 Honest edge: if the agent **fetches more input by itself
-with tools**, the boundary blurs — an outside judgment held that "the store is never used in full" and
+definition, step outside their own input.
+
+🟥 **Honest edge**: if the agent **fetches more input by itself with tools**, the boundary blurs — an outside judgment held that "the store is never used in full" and
 "the swallowed exception" are catchable by ⓐ and ⓒ as well, since those reviewers grep for themselves.
 Conversely, "a rule another repo retired long ago" **cannot be fetched by any tool** — there is no reason
 to have access to that project's review history in the first place. That is where ⓓ remains.
@@ -554,22 +614,42 @@ Full spec: [`fh_integration_contract.md`](knowledge/shared/harness-core/fh_integ
 
 ---
 
-## The forge
+## The forge — where the name comes from
 
-forge-harness treats a project like steel — and the metaphor is literal, not decoration. Work is shaped,
+forge-harness treats a project like steel, and the metaphor is literal, not decoration. Work is shaped,
 hardened by attack, and only then does it ship faster, for having survived.
 
-| Movement | What happens | The commands |
+> 🟥 **This is the name's origin, not a procedure — do not count it.** The smith's words below are a
+> vocabulary, not a stage list. They are a **different layer** from the three-stage process, the four
+> engines, the four-axis commit gate and the six verification axes above; nothing here lines up with a
+> number in any of those, and reading it as a fifth numbered set is the one mistake this section can
+> cause.
+
+**Three at the anvil:**
+
+| The smith's word | What it means here | The commands |
 |---|---|---|
 | **Forge** | shape the raw project into a harness — raise its floor | `install-wizard`, "harness-ify this project" |
 | **Quench** | harden it by attack — the cold pass leaves standing only what is sound | `steel-quench` · `phantom-quench` |
 | **Temper** | take the brittleness back out of the hardened asset | `steel-quench` Wave-T · `templates/temper_check.sh` |
-| → **Accelerate** | a blade that survived the forge cuts faster | `goal-quench` — *Pass → Accelerate* |
 
-All four movements ship. Temper was named before it was built — deliberately (see
-[`ETHOS.md`](docs/ETHOS.md#the-forge)) — and shipped once measurement runs validated it. Around the forge,
-two more signatures keep it running: `harvest-loop` (each session's lessons become permanent skills) and
-`agent-composer` (orchestrate the dispatch). The other skills wait until you need them — full list below.
+```mermaid
+flowchart LR
+  F["Forge"] --> Q["Quench"] --> T["Temper"] --> A(["⟹ Accelerate"])
+  style A fill:#0f766e,stroke:#0f766e,color:#fff
+```
+
+**⟹ And then it accelerates.** A blade that survived the forge cuts faster — `goal-quench`,
+*Pass → Accelerate*. Speed is what the three above **produce**; it is not a fourth thing you do.
+That is the same sentence as the tagline at the top of this page, said in the smith's words:
+quality is the lever, speed is the result.
+
+Every command named above ships today. Temper was named before it was built — deliberately (see
+[`ETHOS.md`](docs/ETHOS.md#the-forge)) — and shipped once measurement runs validated it.
+
+Around the forge, two more signatures keep it running: `harvest-loop` (each session's lessons become
+permanent skills) and `agent-composer` (orchestrate the dispatch). The other skills wait until you need
+them — full list below.
 
 ## 40 skills · 8 agents
 

package/README.zh.md

@@ -18,14 +18,19 @@
 </p>
 
 <p align="center">
-  <b>有需要就交给它。同一个请求反复出现时，它会主动提议把那件事本身做出来。</b>
+  <b>这是一个把质量门禁内置进去的元框架（meta-harness）。</b>
 </p>
 
 <p align="center">
-  你大概已经在对 Claude Code 反复说同样的话：要跑的检查、要守的规则、一次变更该有的样子。<br>
-  <b>forge-harness 把这些变成可复用的东西</b>：住在你仓库里、会自己触发的技能、闸门和 agent。<br>
-  技能刻意保持通用，在使用过程中按你的场景当场锻造。
-  <b>当同一种形状反复回来，它就提议出货</b> —— 作为独立技能，或独立框架。
+  项目、技能、框架 —— 造出来、验过、再加速：这些都在这里交代。<br>
+  它不会就这么把结果递还给你，而是先让这份工作穿过好几道会以 <i>不同方式</i> 失败的检查。<br>
+  <b>而当同一个请求反复回来，它就替你造出那个专门干这件事的框架。</b>
+</p>
+
+<p align="center">
+  你大概已经在对 Claude Code 反复说同样的话：要跑的检查、要守的规则、一次变更该有的样子。
+  变得可复用的正是这一部分，而它刻意保持通用的形态，好在使用过程中按你的场景当场锻造。<br>
+  <sub>长起来的是尝试的次数：试错从你身上剥离，并行地跑。</sub>
 </p>
 
 ---
@@ -39,22 +44,38 @@ git clone https://github.com/chrono-meta/forge-harness.git ~/projects/forge-harn
 cd ~/projects/forge-harness && claude
 ```
 
-**然后输入 `hi`。** 会出现一个带编号的菜单，之后由工具引导你：选一个入口，回答几个问题，
-它会替你运行安装向导。这条线以下是需要时再查的参考资料，而不是开始前的作业。
+**然后输入 `你好`** —— 或者用你真正在用的那门语言打招呼：`hi`、`안녕`、`こんにちは`、`hola`、
+`bonjour`。**不管你用哪种语言问候，菜单都会打开**，并且它会尝试用那种语言回复。会出现一个带编号
+的菜单，之后由工具引导你：选一个入口，回答几个问题，它会替你运行安装向导。
+
+<sub>🟥 <b>关于最后这一点，说实话</b>：语言对齐是一条 <b>背后没有任何机械兜底</b> 的散文规则，所以
+它并不总是成立。2026-08-21 在 floor 档做的盲测，跑在 <b>一个跟你刚建的那份一样的干净克隆</b> 上：
+中文和韩文的问候都完整地用同一种语言回来了，连门的标签都翻了。而在维护者自己那台
+<b>把默认语言钉死了的机器</b> 上，中文 <b>5 次里只中了 1 次</b> —— 这条注脚之所以存在，就是因为它。
+现在仍然不稳的是 <b>菜单会不会弹出来</b>：有一种问候写法就没能唤出菜单。如果它回错了语言、或者
+没给你菜单，直接说一声，它就会切过来。这条残留被如实写在 <code>CLAUDE.md</code> §Voice/Tone 里，
+而不是被抹平。</sub>
+
+这条线以下是需要时再查的参考资料，而不是开始前的作业。
 
-**它放大什么**：尝试的次数。试错从你身上移开，并行运行。
-**它不放大什么**：模型的天花板。框架只把模型抬到它自己的天花板，不会更高。
-**如何验证**：它公开自己的评级。五个身份，如实记录在每一次
-[发布](https://github.com/chrono-meta/forge-harness/releases)中。未变绿的那些，会指名说出还缺哪一次真实运行。
+- **它放大什么** —— 尝试的次数。试错从你身上移开，并行运行。
+
+- **它不放大什么** —— 模型的天花板。框架只把模型抬到它自己的天花板，不会再往上推。
+
+- **如何验证** —— 它公开自己的评级。它把 **自己声称是什么** 名列为五项（框架集群 · 项目孵化器 ·
+  治理门禁 · 前沿 → 组织传导 · 放大器），并在每一次
+  [发布](https://github.com/chrono-meta/forge-harness/releases)里逐项如实打分。未变绿的那些，
+  会指名说出还缺哪一次真实运行。<br>
+  <sub>这五项各自是什么，在下面的《五重身份》一节里展开。</sub>
 
 ---
 
 <p align="center">
-  <img src="docs/pillars.svg" alt="FORK - ADAPT - COLLABORATE - EMPOWER" width="680">
+  <img src="docs/pillars.svg" alt="HARNESS - FORGE - ACCELERATE - COMPOUND" width="680">
 </p>
 
 <p align="center">
-  <b>质量是杠杆，速度是结果。</b> <i>Fork 它。改名。让它成为你的。</i><br>
+  <b>质量是杠杆，速度是结果。</b><br>
   <sub>如果这对你有用，⭐ 一下能帮助更多人发现它。</sub>
 </p>
 
@@ -115,7 +136,9 @@ cd ~/projects/forge-harness
 claude
 ```
 
-> ✅ 然后 **打一句招呼（"hi"）** —— 🐿️ 门菜单是在你打出招呼时出现的，光是启动不会出现。
+> ✅ 然后 **打一句招呼** —— `你好`、`hi`、`안녕`、`こんにちは`，用你习惯的语言就行，它会尝试用
+> 那种语言回复（见上面的注记 —— 多数时候对，但不是每次）。
+> 🐿️ 门菜单是在你 *打出* 招呼时出现的，光是启动不会出现。
 > 说 **"连接一个项目"** → 中枢扫描 `../`，找到 `.git` 目录，创建 `tracks/{project}/`。
 > 想做完整的初始设置（hooks · 门禁 · 基线 —— 每一项单独批准，拒绝会被尊重并记录），
 > 请要 **`/install-wizard`**。
@@ -124,8 +147,8 @@ claude
 
 **你的头 15 分钟** —— 成功长什么样，以及拿它做什么：
 
-1. 当一句招呼（"hi"）能让 🐿️ 门菜单出现、而"连接一个项目"能建出 `tracks/{your-project}/` 时，
-   你就知道设置成功了。
+1. 当一句招呼（任何语言都可以）能让 🐿️ 门菜单出现、而"连接一个项目"能建出
+   `tracks/{your-project}/` 时，你就知道设置成功了。
 2. 然后在同一个会话里拿下一个即时收益：说 **"加速这个项目"**（一份值得接线的技能/插件排序方案，
    安装要过门禁），或者 **"跑一下 /context-doctor"**（token 浪费扫描）。
 3. 一条诚实说明：FH 的核心回报是 **复利累积** —— 会话记录、收割来的学习、跨会话记忆。它从
@@ -296,11 +319,24 @@ Project B  ──→  在 CLAUDE.md 中连接中枢
 
 **三段工序** —— 这是一个 *投入的顺序*，不是一份菜单：
 
+```mermaid
+flowchart LR
+  S["① 立坐标系<br/>设计之前"]
+  P["② 并行去相关<br/>中段"]
+  B["③ 烧一遍<br/>六条轴上"]
+  A(["⟹ 于是它快起来"])
+  S --> P --> B --> A
+  style A fill:#0f766e,stroke:#0f766e,color:#fff
+```
+
+**速度是末尾那支箭，不是第四个方框。** 而 ② 是 **两个分开拧的旋钮** —— *去相关*（应对盲点风险：
+换一个模型家族 ⓐ、换一个立场 ⓑ）与 *并行*（应对表面大小：一个上下文装得下吗）。别做乘法，要挑。
+
 ```
 ① 设计之前先立坐标系     判断坐标系「最先」进场 —— 成功 · 偏向 · 不在范围 · 绝不做 ——
                         而不是事后补写成一份"我做了什么"的记录
 
-② 中段做去相关，用来加速  把工作拆成会以「不同方式」失败的检查，然后一次性跑掉。要挑「哪些
+② 并行去相关              把工作拆成会以「不同方式」失败的检查，然后一次性跑掉。要挑「哪些
                         差异算数」—— 再来一位同一种类的审阅者不是去相关，那是把同一个盲点
                         看两遍。并行本身没有方向，挑方向的是 ① 里那套判断坐标系。
                         这是一种「工作方式」，不是 ③ 里那道收尾检查。
@@ -310,6 +346,11 @@ Project B  ──→  在 CLAUDE.md 中连接中枢
                         看不见的东西照样看不见
 ```
 
+> 📖 **从这里往下，是给想再深入一点的人看的 —— 开始用它并不需要这些。**
+> 如果你是来把它装上、直接开跑的，最上面那个「两分钟」小节就是全部；你可以在这里停住，等哪天
+> 某道检查让你觉得意外了再回来。下面讲的是这些门禁为什么长成这样，是写给 **已经在跑这套中枢的人**
+> 看的。碰到不认识的词 → [`GLOSSARY.md`](knowledge/shared/GLOSSARY.md)。
+
 **六条验证轴** —— 所谓"我们审过了"，往往到头来只做了其中第一条。
 
 🟥 **轴不是按「有多对抗」来分的，是按「它拿到了什么」来分的。** 拿到的东西一样，你堆多少位审阅者，
@@ -317,13 +358,39 @@ Project B  ──→  在 CLAUDE.md 中连接中枢
 
 | 轴 | **它拿到什么** | 它抓到什么 | 典型手段 |
 |---|---|---|---|
-| **ⓐ 不同家族** | diff + 作者的框架叙述 | **实现** 错了 | 换一个模型家族的审阅者（`auto-decorrelation`） |
-| **ⓑ 立场 (standpoint)** | diff + **目标框架自己的正典** | **你引用的那条规约是不是真这么说** | 在那个框架自己的仓库与规则里去跑这份 diff ([`§7`](knowledge/shared/harness-core/field_verdict_crossfamily_gate.md)) |
-| **ⓒ 隔离接地** | 作者写下的那些句子 + 此刻的这棵树 | **主张** 错了 | 找一个没写过它的人，把它说的重新测一遍 |
+| **ⓐ 不同家族** | 变更 + 作者的框架叙述 | **实现** 错了 | 换一个模型家族的审阅者（`auto-decorrelation`） |
+| **ⓑ 立场 (standpoint)** | 变更 + **目标框架自己的正典** | **你引用的那条规约是不是真这么说** | 在那个框架自己的仓库与规则里去跑这份变更 ([`§7`](knowledge/shared/harness-core/field_verdict_crossfamily_gate.md)) |
+| **ⓒ 隔离接地** | 作者写下的那些句子 —— 他的主张，*以及* 他在 **动手之前声明过的东西** —— + 此刻的这棵树 | **主张** 错了 · 增量与当初声明的对不上 | 找一个没写过它的人，把它说的重新测一遍。至于事前声明，那是一道把写下的成功定义拿回来对照增量的门禁 |
 | **ⓓ 第三方对面** | 问题 + **别人的代码库** | **这是不是早就被解过了** · 你的变更会碰到别人仓库的哪里 | 在一个无关的第三方仓库里看同一个问题 |
 | **ⓔ 首次真实使用** | 一个实打实的目标 | **你测量的方式** 错了 —— 量具的量具 | 拿一个真实目标真跑一次，然后动手核对那份结果 |
 | **ⓕ 撤回并观察** | 把接线删掉之后的那棵树 | **锚** 错了 —— 那道检查是装饰 | 把它所守护的东西删掉，确认 *正是那一条* 检查变红 |
 
+> **ⓒ 在 2026-08-21 被拓宽了，而「怎么拓宽的」才是更有用的那一半。** 本仓库的提交标记自
+> 2026-08-09 起就要求写上作者自己的 **事前声明** —— *什么算成功* 与 *什么绝不做* —— 而且必须写在
+> 设计 *之前*。那天带着对照组一测：**没有任何代码在读它。** 消费它的代码整整零行，而兄弟字段在
+> 21 处被检查；门禁规格甚至没有点过它的名。在真实语料里，**98 份标记中有 37 份根本没有那一行**
+> —— 其中还包括一份跑过 28 条 lane、别的字段全部填满、并通过了评审面板的标记。这些轴全都朝
+> *外* 看 —— 变更、目标仓库、先例、产物。**没有一条轴回头看那份记录自己的必填字段。**
+> 一个没有消费方的槽位永远报告「完成」，**因为是「存在」本身在替你做判定。**
+>
+> 修法不是加第七条轴。ⓒ 本来就拿到 *作者写下的那些句子 + 此刻的这棵树*，而这与一次事前声明检查
+> 所拿到的东西 **逐字相同**。时态（先声明的，还是事后主张的）和对抗性一样，是一种 **姿态**，
+> 不是一条轴。真去新铸一条，就等于把上面那次盲判重分类抓到的错误再犯一遍。
+
+> **同侪会话记在 ⓑ 和 ⓓ 上 —— 2026-08-21 决定，而「问哪一个同侪」才是全部诀窍。** 同一个框架分叉
+> 出去、正在另一条支线上跑得火热的并行会话，不是你的副本。在它跑热的那一点上，它确实长出了
+> **第二张脸**：一个货真价实的立场（ⓑ），一个货真价实的「别人的代码库」（ⓓ）。所以同侪的判定记在
+> 这两条轴上 —— 没有新铸第七条轴，标记里的 `axes-run` 字母表也没有变。
+>
+> 上面那张表的 ⓓ 行之所以写的是「**别人的代码库**」而不是「别家公司的仓库」，原因就在这里：真正
+> 起作用的边界是 **那份判定是在谁的工作脉络里产生的**，而不是这些文件挂在谁的 GitHub 组织名下。
+> 一个在另一条支线上跑热的同侪，在这条边界 **之外**；而你从当前这个脉络里派发出去的子 agent，
+> 不管它去读的是多么不相干的仓库，都仍然在这条边界 **之内**。
+>
+> 🟥 咬人的是那条推论。**要问同侪的，是它真正跑热的那条轴。** 出了这个范围，它戴的就是你的脸，
+> 问了也去不了相关 —— 同样的输入，同样的盲点。而且子 agent 顶替不了它，重读一遍自己的文字也
+> 顶替不了：那第二张脸来自那个会话真的 *做了* 不一样的事，靠提示词是造不出来的。
+
 **你不必每次都把六条跑满，这正是设计** —— 别做乘法，要 **挑**：
 
 ```
@@ -343,7 +410,9 @@ Project B  ──→  在 CLAUDE.md 中连接中枢
 
 **为什么这一套不会被基础模型的进步取代** —— 一条轴是由 **输入** 定义的，不是由 *审阅者的能力*
 定义的。模型再强，**「没拿到的信息」它依然看不见。** 脚手架会随着模型变好而脱落，但 **输入边界上的
-去相关不会脱落**，而单个作者按定义就走不出自己的输入。🟥 诚实的边缘地带：如果 agent **自己用工具
+去相关不会脱落**，而单个作者按定义就走不出自己的输入。
+
+🟥 **诚实的边缘地带**：如果 agent **自己用工具
 去取更多输入**，这条边界就会变模糊 —— 外部判定确实认为「整个存储从未被完整使用」和「异常被吞掉」
 这两条 ⓐ · ⓒ 也能抓到（因为它们会自己 grep）。反过来，「别人的仓库当年废弃掉的某条规则」
 **用工具也取不到** —— 你压根没有理由去访问那个项目的评审历史。ⓓ 留下来的位置就在那里。
@@ -461,21 +530,37 @@ hooks 不会自动触发，M2 的 agent 派发步骤需要适配器（或交互
 
 ---
 
-## 大锻炉 (The forge)
+## 大锻炉 (The forge) —— 名字的由来
 
-forge-harness 把项目当作钢来对待 —— 而这个隐喻是字面的，不是装饰。工作被塑形、以攻击淬硬，
+forge-harness 把项目当作钢来对待，而这个隐喻是字面的，不是装饰。工作被塑形、以攻击淬硬，
 唯有如此才更快出炉，因为它挺过了考验。
 
-| 工序 | 发生了什么 | 命令 |
+> 🟥 **这是名字的由来，不是流程 —— 不要数它。** 下面那些铁匠的用词是一套 **词汇**，不是阶段清单。
+> 它与上文的 **三段工序 · 四大引擎 · 四轴门禁 · 六轴验证是不同的层**，不对应其中任何一个数字；
+> 把它读成「第五个带编号的集合」，是本节唯一可能引起的误会，所以在这里先掐断。
+
+**铁砧上的三道：**
+
+| 铁匠的用词 | 在这里是什么意思 | 命令 |
 |---|---|---|
 | **锻造 (Forge)** | 把生坯项目塑形为框架 —— 抬高其下限 | `install-wizard`、"把这个项目框架化" |
 | **淬火 (Quench)** | 以攻击将其淬硬 —— 冷审阅只让健全的东西留存 | `steel-quench` · `phantom-quench` |
 | **回火 (Temper)** | 把淬硬资产里的脆性 (brittleness) 再退掉 | `steel-quench` Wave-T · `templates/temper_check.sh` |
-| → **加速 (Accelerate)** | 一把挺过锻炉的刀刃切得更快 | `goal-quench` —— *Pass → Accelerate* |
 
-四道工序全部出货。回火 (Temper) 在被造出 *之前* 就先起了名字 —— 刻意为之（见
-[`ETHOS.md`](docs/ETHOS.md#the-forge)）—— 并在测量运行验证之后出货。围绕这座锻炉，还有两个
-签名部件让它持续运转：`harvest-loop`（每次会话的教训成为永久技能）与
+```mermaid
+flowchart LR
+  F["锻造 Forge"] --> Q["淬火 Quench"] --> T["回火 Temper"] --> A(["⟹ 加速 Accelerate"])
+  style A fill:#0f766e,stroke:#0f766e,color:#fff
+```
+
+**⟹ 然后，它自己就快起来了。** 一把挺过锻炉的刀刃切得更快 —— `goal-quench`，*Pass → Accelerate*。
+速度是上面这三道 **锻出来的东西**，不是你另外再做的第四件事。这和本页最上面那句标语说的是同一件事，
+只不过换成了铁匠的用词：质量是杠杆，速度是结果。
+
+上面点到名的命令，如今全部已经出货。回火 (Temper) 在被造出 *之前* 就先起了名字 —— 刻意为之
+（见 [`ETHOS.md`](docs/ETHOS.md#the-forge)）—— 并在测量运行验证之后出货。
+
+围绕这座锻炉，还有两个签名部件让它持续运转：`harvest-loop`（每次会话的教训成为永久技能）与
 `agent-composer`（编排派发）。其余技能等你需要时再出场 —— 完整清单见下。
 
 ## 40 skills · 8 agents

package/docs/ETHOS.md

@@ -18,14 +18,21 @@ Everything below is **copyable**. None of it is a secret. The principles are the
 FH treats a project like steel — heat, shape, and shock, in named movements. The metaphor is literal,
 not decoration:
 
-| Movement | What it does | Today |
+**Three at the anvil:**
+
+| The smith's word | What it does | Today |
 |---|---|---|
 | **Forge** | shape the raw project into a harness — raise its floor | `install-wizard`, harness-ify |
 | **Quench** | harden it by attack — the cold pass leaves standing only what is sound | `steel-quench`, `phantom-quench` |
 | **Temper** | take the brittleness back out of the hardened asset | `steel-quench` **Wave-T** + `templates/temper_check.sh` |
-| → **Accelerate** | a blade that survived the forge cuts faster — pass, then run | `goal-quench` · *Pass → Accelerate* |
 
-All four movements now ship. **Temper** spent its first months named-but-unbuilt — deliberately, per
+**⟹ And then it accelerates.** A blade that survived the forge cuts faster — `goal-quench`,
+*Pass → Accelerate*. Three at the anvil, and speed is what those three **produce**; it is not a
+fourth thing you do at the anvil. Quality is the lever, speed is the result.
+
+Every command named above ships today. (These are the smith's words — a **vocabulary**, not a stage
+list, and a different layer from the three-stage process, the four engines, the four-axis commit gate
+and the six verification axes. Do not count them against those.) **Temper** spent its first months named-but-unbuilt — deliberately, per
 principle 5 — and shipped only after measurement runs on independent quench convergences validated that
 the check flags over-hardening without punishing simplification. Quenched steel is hard but brittle;
 no smith ships it un-tempered, and now neither does FH: after convergence, Wave-T measures the complexity

package/knowledge/shared/dialogue/ai_dialogue_playbook.md

@@ -91,6 +91,109 @@ API 키 대화창 기록 방지 — Write 툴로 `.env` 직접 생성 (대화창
 
 **Forbidden response**: "I can't do that — I'm not in that project's cwd." Always check if Agent dispatch covers it first.
 
+🟥 **A live peer session is a different thing from a subagent — and the default reading is wrong.**
+Operator, 2026-08-21, correcting exactly that: *"이작업을 분할하라는게 아니라 논의를 해보라는
+뜻이었어 작업은 니가하고"* → *"주위피어에는 **두뇌를빌려서 네 수행에 대한 판단**을 도움받아보라는
+거였으니"*.
+
+| | you dispatch it | you ask a peer |
+|---|---|---|
+| what moves | **the work** | **the judgment** |
+| what comes back | a completed unit | a verdict on *your* execution |
+| the peer's own work | — | keeps running; you did not take it |
+
+The reflex is to read "talk to the peers" as "split the task across the peers", because a peer
+looks like a bigger subagent. It is not. The operator's own framing, same day:
+
+> *"예열되어 다른갈래로 뻗어있는 자신의 다른 면모들은 하나의 **별개하네스(동일맥락을 공유하는)**
+> 로도 동작도가능하므로 인사이트를 얻을수있다"*
+
+A peer is **pre-warmed on its own branch of the same context** — a separate harness that shares
+your origin but not your recent path. The operator's sharper form, same day:
+
+> *"동일 하네스에서 분화한 세션이라도 **그 세션으로 뜨거워진 지점에서는 다른 양상과 관점을
+> 지니게 된다. 그 세션만큼은 두 얼굴이 생기는거다.**"*
+>
+> *"**자신에 대해서는 뜨겁게, 남에 대해서는 그만큼 차갑게** (동일 하네스라도)"*
+
+The divergence is not in how context is *delivered* — the peer genuinely **holds a different
+aspect** at the point it got hot. The second line names the mechanism, and it is the half that
+is easy to miss: the same heat that makes a session sharp on **its own** thread makes it
+correspondingly **cold on yours**. Both halves are load-bearing and they buy different things:
+
+| | what it buys |
+|---|---|
+| **hot on its axis** | it sees what you structurally cannot — it did that work |
+| **cold on yours** | it does not carry your optimism, your sunk cost, or your reading of your own claim |
+
+The second is why "read it again, more carefully" never substitutes: **you cannot make yourself
+cold about your own output.** Dividing the work destroys both properties at once — a peer given
+part of your task becomes hot on it, and is then judging its own output.
+
+🟥 **Two operating consequences, which follow from the mechanism rather than from etiquette:**
+- **Ask the peer about the axis it got hot on.** Off that axis a peer wears *your* face — asking
+  buys nothing and costs a round trip.
+- **A subagent or your own re-read cannot substitute.** The second face comes from that session
+  having actually done the work; a prompt cannot manufacture heat, and nothing manufactures
+  coldness toward your own output.
+
+**Measured, and it is not one good day.** Scanned this hub's own records for peer-attributed
+catches: **119 mentions across 35 files spanning 13 distinct dates (2026-08-08 → 08-21)**;
+cross-family sidecar catches count separately at **113** — comparable in size, a *different*
+axis. Eight sampled hits were hand-verified as genuine peer attribution, and at least two run
+the **other way** (a peer's misdiagnosis reversed by measurement here) — so the shape is mutual
+contention, not one-way review.
+⚠️ Named residuals: these are *mentions*, not deduplicated incidents (one event can be described
+in several files); attribution is self-reported in this hub's own records; the window is the
+~2-week parallel-session era only.
+
+🟥 **Whether this is a verification axis in its own right is an OPEN OPERATOR DECISION, not
+settled here.** The operator also said *"이것도 6축검증법에 속한다"*, which places peer discussion
+inside the canonical six axes — but which axis, or whether it is a seventh, is unresolved: it is
+closest to ⓒ isolated-grounding (someone re-verifies *your record*), yet a peer is precisely
+**not** isolated. Do not cite this paragraph as if the axis question were closed.
+See `tracks/_meta/fh_signal_2026-08-21_peer-axis-canon-claim.md`.
+
+**So the ask to a peer is a question, not a work item.** The operator's own modelling of the
+boundary, same day, on a branch belonging to another session: *"그 브랜치는 08d6fe75 갈래
+것이고 **여기서 손댈 게 아니다. 판단만 낸다.**"* — when the artifact belongs to another
+session, produce a verdict, never an edit. This is about **scope of authority** and is separate
+from the shared-checkout rules (which are about not clobbering); both apply at once.
+
+### The effect that needs 2+ concurrent — and why it is not a preference
+
+Operator, 2026-08-21, arriving at it against their own habit:
+
+> *"내가 두개이상을 동시에 잘 쓰려고는 안하는데 아이러니하게도 **두개이상을 동시에 써야 얻을수
+> 있는 효과**도있어보이네"* → *"**2명~3명이 하네스를 굴려서 경합시키는 모양새**가 나오니까"*
+
+The shape is not "parallel work". It is **contention** — two or three sessions *driving the same
+harness* and colliding. What that produces cannot be reached serially, because both halves of the
+hot/cold table must exist **at the same time**: one session hot on a thread while another is cold
+on it. Run them in sequence and you get one face twice.
+
+Same day, in both directions: the wiring-hot session found a defect in this session's delta; this
+session found a drift defect inside *that* session's delta — and the check that caught it fires on
+*"did main change today"*, not *"did I change it"*, which is a contention-shaped trigger by
+construction. Neither finding was self-caught by its author.
+
+⚠️ **An observed effect, not a promotion to default.** Concurrency costs coordination, and it is
+where the same day's two shared-checkout accidents came from. The claim is narrow: *some* effects
+are unobtainable without it.
+
+**So how to hold it** (operator, 2026-08-21): *"디폴트 승격은 안 해도 되. 다만 병렬세션 띄워서
+운영자가 작업할 때 얻게 되는 **반사이익**이라고 보면 되지. 아니면 **네가 권유해도 괜찮음**.
+「병렬세션으로 나눠서 작업해도 좋을 것 같습니다」 이런 식으로"*.
+
+Two things follow, and the second is a permission most sessions will not assume they have:
+- **It is a by-product, not a reason.** The operator opens parallel sessions for their own reasons;
+  contention is what they *also* get. Do not argue for concurrency on the strength of this effect
+  alone.
+- 🟥 **You may propose it.** When work has separable threads and a second face would help — a
+  decision you would otherwise ratify alone, a claim only its author has read — say so plainly:
+  *"병렬세션으로 나눠서 작업해도 좋을 것 같습니다."* Proposing is invited; deciding is not yours,
+  and the standing preference remains delegation-in-conversation over standing parallelism.
+
 **Context Card** (required for non-trivial dispatch):
 ```
 [Session Context Card]
@@ -120,6 +223,34 @@ Note: {constraints the agent must know}
 
 ---
 
+## Approval Must Be Evaluable — never hand over a blind stamp
+
+Operator, 2026-08-21: *"내가 판단해야하는건 내가 알아볼수있게 쉽게 브리핑해줘 **승인하더라도
+블라인드 도장은 찍고싶지않아**."*
+
+An approval request is not done when it is *complete*; it is done when the approver can **reach
+the verdict themselves**. A dump of everything measured technically discloses more and decides
+less — the approver ends up ratifying your conclusion rather than forming one.
+
+**What a decision-ready item carries:**
+```
+the decision       stated as a choice, not as a status report
+what it costs      what becomes irreversible, what stays open
+the evidence       the number AND how it was measured (an uncalibrated number is not evidence)
+your recommendation  named as yours, so it can be rejected without re-deriving everything
+what you did NOT check  the residual, by name — this is the half that makes the rest trustworthy
+```
+
+🟥 **The failure mode is not withholding — it is volume.** Burying the one load-bearing fact in a
+complete record is the same defect as omitting it: either way the approver cannot separate what
+matters. If it takes the approver a re-derivation to decide, the briefing has handed them a blind
+stamp with extra steps.
+
+**Applies to every approval surface**, not only long briefs: a one-line "merge this?" still owes
+the choice, the cost, and the residual.
+
+---
+
 ## Counter-Argument Protocol
 
 When the user pushes back on an AI recommendation ("is that right?", "something seems off"):