Poofread blogs and make minor modifications

.github/workflows/pages.yml

@@ -4,9 +4,6 @@ on:
   push:
     branches:
       - main
-    paths:
-      - site/**
-      - .github/workflows/pages.yml
   workflow_dispatch:
 
 permissions:
@@ -28,15 +25,29 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          cache: npm
+          cache-dependency-path: site/package-lock.json
+
+      - name: Install site dependencies
+        run: npm ci
+        working-directory: site
+
+      - name: Build site
+        run: npm run build
+        working-directory: site
+
       - name: Configure Pages
         uses: actions/configure-pages@v5
 
       - name: Upload site artifact
         uses: actions/upload-pages-artifact@v3
         with:
-          path: site
+          path: site/dist
 
       - name: Deploy to GitHub Pages
         id: deployment
         uses: actions/deploy-pages@v4
-

.gitignore

@@ -1,3 +1,4 @@
 .DS_Store
 export
 node_modules
+site/dist

README.md

@@ -2,73 +2,126 @@
 
 Languages: [English](README.md) | [Chinese](README.zh.md)
 
-GDPevo is a public benchmark for evaluating self-evolving agents on
-economically valuable, real business tasks. To our knowledge, it is the first
-GDP-valued benchmark that treats agent evaluation as a stateful process: an
-agent first works through related train tasks, turns experience into reusable
-skills, and is then evaluated on held-out tasks from the same business
-environment.
+**GDPevo** is a public benchmark for measuring agent self-evolution on real
+business work.
 
-Most agent benchmarks still evaluate stateless task completion. GDPevo instead
-asks whether agents can improve through experience: can they learn business
-rules, source precedence, operating procedures, and output discipline from
-earlier work, and can that learning make later work more accurate and cheaper
-to execute?
+The question is not only: *can the agent solve this task?*
 
-The benchmark can be used to evaluate:
+It is: **can the agent evolve across related tasks and get better on held-out
+tasks from the same business world?**
+
+The first public release contains **120 GDP-worthy tasks** across **12 task
+groups** in CRM, ERP, and Finance. Each task group contains one shared business
+environment, **5 train tasks**, and **5 held-out test tasks**. Train tasks drive
+the evolve step; test tasks measure whether the resulting update transfers to
+later work it has not seen.
+
+## What GDPevo Measures
+
+Most agent benchmarks measure stateless task completion. GDPevo measures a
+stateful loop:
+
+1. An agent works through related train tasks.
+2. It turns train-task evidence into reusable procedures, memory, or operating rules.
+3. It is evaluated on held-out tasks from the same business environment.
+
+This makes GDPevo useful for evaluating:
 
 - self-evolving or continual-learning agents;
-- skill creators and skill optimizers;
-- end-to-end agent memory systems.
+- evolution-update and self-improvement mechanisms;
+- end-to-end agent memory systems;
+- whether self-evolution improves both **accuracy** and **cost**.
 
-The first release contains 120 tasks organized into 12 task groups. Each task
-group has one shared business environment, five train tasks, and five test
-tasks. The task groups are constructed from economically meaningful industry
-workflows, including finance, enterprise CRM, and ERP automation.
+The benchmark is built around real company interfaces: CRM, ERP, finance,
+procurement, support, lending, investment, HR, and reporting workflows. Task
+groups are seeded from real-job sources such as GDPval and SOP-Bench, then
+expanded into shared environments with lookalike records, hidden operational
+rules, and deterministic rule-based graders.
 
-In the released Codex GPT-5.5 xhigh run, evolved agents improve accuracy by
-18.21 percentage points on average after inductive learning, while reducing
-token cost by 25.75% on average. The released artifacts include executable task
-groups, evaluation reports, generated skill packages, and a reusable evaluation
-workspace that can automate the full scoring flow.
+## Evaluation Setup
 
-## Repository Layout
+Released runs compare three modes:
 
-| Path | Purpose |
+| Mode | Meaning |
 | --- | --- |
-| `data/` | Released task group data, data board, shared environments, train/test tasks, answers, evaluators, and data notes. |
-| `experiments/` | Released evaluation protocol, evaluation workspaces, result reports, generated skills, and experiment board. |
-| `site/` | GitHub Pages scaffold and public site assets. |
-| `assets/` | Figures, logos, and visual assets used by released materials. |
+| `base` | The agent solves the held-out test tasks cold, with no prior exposure to train tasks. |
+| `demo` | The agent reads train tasks with gold answers, distills reusable procedures, then takes the test. |
+| `reflect` | The agent attempts train tasks without answers, receives graded feedback, updates its procedure from mistakes, then takes the test. |
+
+Scores use `avg@3`: mean held-out accuracy over 3 attempts per task.
+
+Cost is reported in **USD** from raw token metrics in the report YAML files.
+The released boards show token counts in thousands (`k`) and cost rounded to
+two decimals.
+
+## Released Results
+
+Across all three harnesses, the same pattern holds: self-evolution improves
+held-out accuracy by roughly **+17 to +22 percentage points**.
+
+| Harness | Model | Thinking level | `base` avg@3 | `demo` avg@3 | `reflect` avg@3 | Accuracy lift | Cost change |
+| --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
+| Codex | GPT-5.5 | xhigh | 48.35% | 65.99% | 67.13% | +18.21 pp | -25.75% |
+| Claude Code | Opus 4.8 | xhigh | 49.11% | 70.90% | 67.94% | +20.31 pp | -8.69% |
+| Panofy | Opus 4.6 | high | 50.17% | 68.24% | 67.98% | +17.94 pp | +11.82% |
+
+Full per-task reports are under:
+
+- [`experiments/codex_gpt5_5_xhigh/`](experiments/codex_gpt5_5_xhigh/)
+- [`experiments/claude_code_opus_4_8_xhigh/`](experiments/claude_code_opus_4_8_xhigh/)
+- [`experiments/panofy_claude_opus_4_6_high/`](experiments/panofy_claude_opus_4_6_high/)
+
+The aggregate board is
+[`experiments/EXPERIMENT_BOARD.md`](experiments/EXPERIMENT_BOARD.md).
 
 ## Data
 
-Each task group contains one shared business environment, five train tasks, and
-five test tasks. Train tasks provide the experience source, and test tasks
-measure whether the resulting skills improve later work in the same business
-environment.
+Each task group contains:
 
-Released task groups are summarized in [data/DATA_BOARD.md](data/DATA_BOARD.md).
-See [data/README.md](data/README.md) for the data layout and task group format.
+- one shared business environment;
+- 5 train tasks;
+- 5 held-out test tasks;
+- answer files and deterministic evaluators;
+- notes describing the data and grading setup.
 
-## Experiments
+Released task groups are summarized in
+[`data/DATA_BOARD.md`](data/DATA_BOARD.md). See
+[`data/README.md`](data/README.md) for the data layout and task group format.
 
-The released evaluation run compares three conditions:
+## Evaluation Workspaces
+
+Reusable evaluation workspaces live under
+[`experiments/eval_workspace/`](experiments/eval_workspace/):
+
+| Workspace | Purpose |
+| --- | --- |
+| `codex/` | Codex evaluation workflow and guides |
+| `claude_code/` | Claude Code evaluation workflow and guides |
+| `panofy/` | Panofy evaluation workflow and guides |
+| `codex_zh/`, `claude_code_zh/`, `panofy_zh/` | Chinese mirrors |
 
-- `no_skill`
-- `demonstration_skill`
-- `reflection_skill`
+Each workspace describes how to load its guides, run the full evaluation for a
+task group, generate mode-specific artifacts when needed, aggregate `avg@3`, record
+token/cost metrics, and write `report/<task_group_id>.yaml`.
+
+## Repository Layout
+
+| Path | Purpose |
+| --- | --- |
+| `data/` | Released task group data, data board, shared environments, train/test tasks, answers, evaluators, and notes. |
+| `experiments/` | Released evaluation protocols, reusable workspaces, report YAMLs, generated artifacts, and experiment board. |
+| `site/` | React/Vite public site and GitHub Pages build pipeline. |
 
-Results are summarized in
-[experiments/EXPERIMENT_BOARD.md](experiments/EXPERIMENT_BOARD.md).
+## Public Site
 
-See [experiments/README.md](experiments/README.md) for details.
+The public landing page and blog live in [`site/`](site/). The site is built
+with React/Vite and deployed to GitHub Pages.
 
-## Evaluation Workspace
+Local preview:
 
-The reusable evaluation workspace is available at
-[experiments/eval_workspace/](experiments/eval_workspace/). A Chinese mirror is
-available at [experiments/eval_workspace_zh/](experiments/eval_workspace_zh/).
-The workspace describes how Codex can run the full evaluation workflow with
-clean-context skill-generation and solver agents, aggregate `avg@3`, record
-token and cost metrics, and write the final report.
+```bash
+cd site
+npm ci
+npm run build
+npm run preview
+```

README.zh.md

@@ -2,49 +2,123 @@
 
 语言：[English](README.md) | [中文](README.zh.md)
 
-GDPevo 是一个面向真实业务场景的公开 agent benchmark，用于评估自进化 agent 在具有经济价值的任务中学习和迁移规则的能力。据我们所知，它是首个面向 GDP-valued tasks 的 stateful benchmark：agent 先完成一组相关 train tasks，将经验沉淀为可复用 skill，再在同一业务环境下的 held-out test tasks 上接受评估。
+**GDPevo** 是一个公开 benchmark，用来在真实业务工作上衡量 agent 的
+self-evolution 能力。
 
-多数现有 agent benchmark 仍然评估“无状态”的单次任务完成。GDPevo 关注的是另一类能力：agent 能否从早期任务中学习业务规则、信息来源优先级、操作流程和输出规范，并让后续任务在准确率和执行成本上获得提升。
+我们不只问：*这个 agent 能不能解决这个任务？*
 
-这个 benchmark 可以用于：
+而是问：**它能不能借由相关任务完成自我进化，并在同一业务世界的 held-out 任务上变得更好？**
 
-- 评估带有自进化或持续学习能力的 agent。
-- 评估 Skill Creator / SkillOpt 的能力。
-- 评估 Agent Memory 的端到端效果。
+首个公开版本包含 **120 个 GDP-worthy tasks**，覆盖 CRM、ERP 和 Finance
+的 **12 个 task group**。每个 task group 包含一个共享业务环境、**5 个
+train tasks** 和 **5 个 held-out test tasks**。Train tasks 驱动 evolve；
+test tasks 用来衡量这次更新能否迁移到它没见过的后续任务。
 
-首个公开版本包含 120 个任务，并组织为 12 个 task groups。每个 task group 包含一个共享业务环境、5 个 train tasks 和 5 个 test tasks。任务组从具有经济价值的产业工作流构造而来，例如金融、企业 CRM 和 ERP 自动化。
+## GDPevo 衡量什么
 
-在已发布的 Codex GPT-5.5 xhigh 运行中，自进化后的 agent 在归纳学习后平均获得 18.21 个百分点的准确率提升，同时 token 成本平均节省 25.75%。当前公开内容包括可执行 task groups、评估报告、生成的 skill packages，以及可复用的 evaluation workspace，用于自动执行完整评估流程并生成最终分数。
+多数 agent benchmark 衡量的是无状态的单次任务完成。GDPevo 衡量的是一个
+stateful loop：
 
-## 仓库结构
+1. Agent 先处理一组相关 train tasks。
+2. 它把 train-task evidence 转成可复用的流程、memory 或操作规则。
+3. 它再在同一业务环境里的 held-out tasks 上接受评估。
 
-| 路径 | 用途 |
+因此，GDPevo 可以用于评估：
+
+- self-evolving 或 continual-learning agents；
+- evolution update 与自我改进机制；
+- 端到端 agent memory systems；
+- self-evolution 是否同时提升 **accuracy** 并降低 **cost**。
+
+这个 benchmark 围绕真实公司 interface 构建，覆盖 CRM、ERP、finance、
+procurement、support、lending、investment、HR 和 reporting 等工作流。
+Task groups 由 GDPval 和 SOP-Bench 等真实工作来源 seed，再扩展成共享业务环境；
+环境中包含 lookalike records、隐藏业务规则和 deterministic rule-based graders。
+
+## 评估设置
+
+已发布的 runs 比较三种 mode：
+
+| Mode | 含义 |
 | --- | --- |
-| `data/` | 已发布的 task group 数据、数据看板、共享环境、train/test tasks、答案、评测器和数据说明。 |
-| `experiments/` | 已发布的评估协议、evaluation workspaces、结果报告、生成 skills 和实验看板。 |
-| `site/` | GitHub Pages 站点脚手架和公开展示资源。 |
-| `assets/` | 已发布材料使用的图片、logo 和视觉资源。 |
+| `base` | Agent 作为基线完成 held-out test tasks，不接触 train tasks。 |
+| `demo` | Agent 阅读带 gold answers 的 train tasks，提炼可复用流程，再进入 test。 |
+| `reflect` | Agent 先在没有答案的 train tasks 上尝试，收到 grading feedback，从错误中更新做法，再进入 test。 |
+
+分数使用 `avg@3`：每个 task 的 held-out accuracy 在 3 次尝试上的平均值。
+
+Cost 使用 report YAML 中的原始 token metrics 计算，并以 **USD** 汇报。公开看板中
+token 数按千为单位 (`k`) 展示，cost 四舍五入到小数点后两位。
+
+## 已发布结果
+
+三个 harnesses 上都出现了相同趋势：self-evolution 能让 held-out accuracy 提升约
+**+17 到 +22 个百分点**。
+
+| Harness | 模型 | Thinking level | `base` avg@3 | `demo` avg@3 | `reflect` avg@3 | 准确率提升 | Cost 变化 |
+| --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
+| Codex | GPT-5.5 | xhigh | 48.35% | 65.99% | 67.13% | +18.21 pp | -25.75% |
+| Claude Code | Opus 4.8 | xhigh | 49.11% | 70.90% | 67.94% | +20.31 pp | -8.69% |
+| Panofy | Opus 4.6 | high | 50.17% | 68.24% | 67.98% | +17.94 pp | +11.82% |
+
+完整 per-task reports 位于：
+
+- [`experiments/codex_gpt5_5_xhigh/`](experiments/codex_gpt5_5_xhigh/)
+- [`experiments/claude_code_opus_4_8_xhigh/`](experiments/claude_code_opus_4_8_xhigh/)
+- [`experiments/panofy_claude_opus_4_6_high/`](experiments/panofy_claude_opus_4_6_high/)
+
+汇总看板见
+[`experiments/EXPERIMENT_BOARD.zh.md`](experiments/EXPERIMENT_BOARD.zh.md)。
 
 ## 数据
 
-每个 task group 包含一个共享业务环境、5 个 train tasks 和 5 个 test tasks。Train tasks 提供经验来源，test tasks 用于衡量生成的 skills 是否能让同一业务环境下的后续任务获得提升。
+每个 task group 包含：
+
+- 一个共享业务环境；
+- 5 个 train tasks；
+- 5 个 held-out test tasks；
+- 答案文件和 deterministic evaluators；
+- 描述数据和 grading 设置的 notes。
 
-已发布 task groups 汇总在 [data/DATA_BOARD.zh.md](data/DATA_BOARD.zh.md)。
+已发布 task groups 汇总在
+[`data/DATA_BOARD.zh.md`](data/DATA_BOARD.zh.md)。数据目录说明和 task group
+格式见 [`data/README.zh.md`](data/README.zh.md)。
 
-数据目录说明和 task group 格式见 [data/README.zh.md](data/README.zh.md)。
+## Evaluation Workspaces
 
-## 实验
+可复用的 evaluation workspaces 位于
+[`experiments/eval_workspace/`](experiments/eval_workspace/)：
 
-当前公开的评估运行比较三种条件：
+| Workspace | 用途 |
+| --- | --- |
+| `codex/` | Codex evaluation workflow 和 guides |
+| `claude_code/` | Claude Code evaluation workflow 和 guides |
+| `panofy/` | Panofy evaluation workflow 和 guides |
+| `codex_zh/`, `claude_code_zh/`, `panofy_zh/` | 中文镜像 |
 
-- `no_skill`
-- `demonstration_skill`
-- `reflection_skill`
+每个 workspace 都说明了如何加载 guides、为一个 task group 跑完整评估、在需要时生成
+mode-specific artifacts、聚合 `avg@3`、记录 token/cost metrics，并写出
+`report/<task_group_id>.yaml`。
+
+## 仓库结构
+
+| 路径 | 用途 |
+| --- | --- |
+| `data/` | 已发布的 task group 数据、数据看板、共享环境、train/test tasks、答案、评测器和 notes。 |
+| `experiments/` | 已发布的评估协议、可复用 workspaces、report YAMLs、生成 artifacts 和实验看板。 |
+| `site/` | React/Vite public site 和 GitHub Pages build pipeline。 |
+| `assets/` | 已发布材料使用的 figures、logos 和视觉资源。 |
 
-结果汇总在 [experiments/EXPERIMENT_BOARD.zh.md](experiments/EXPERIMENT_BOARD.zh.md)。
+## Public Site
 
-详细说明见 [experiments/README.zh.md](experiments/README.zh.md)。
+公开 landing page 和 blog 位于 [`site/`](site/)。站点使用 React/Vite 构建，并部署到
+GitHub Pages。
 
-## Evaluation Workspace
+本地预览：
 
-中文 evaluation workspace 位于 [experiments/eval_workspace_zh/](experiments/eval_workspace_zh/)，英文版本位于 [experiments/eval_workspace/](experiments/eval_workspace/)。它说明了如何使用 Codex 组织完整评估流程：生成 skill、运行 solver、聚合 `avg@3`、记录 token 和 cost 指标，并写出最终报告。
+```bash
+cd site
+npm ci
+npm run build
+npm run preview
+```

data/DATA_BOARD.md

@@ -4,9 +4,10 @@ Languages: [English](DATA_BOARD.md) | [Chinese](DATA_BOARD.zh.md)
 
 This board summarizes the released GDPevo task groups: 120 tasks across 12
 stateful benchmark units. Each task group has one shared business environment,
-five train tasks, and five test tasks, allowing the benchmark to measure whether
-experience from earlier work improves later work in the same environment. Full
-definitions are available in each task group's `task_group.yaml`.
+five train tasks, and five held-out test tasks, allowing the benchmark to
+measure whether self-evolution from earlier tasks improves later work in the
+same environment. Full definitions are available in each task group's
+`task_group.yaml`.
 
 | task_group_id | scenario_id | domain | focus | train | test | path |
 | --- | --- | --- | --- | ---: | ---: | --- |