docs(eval): commit first real fidelity baseline (59/84 measured, 40% coverage)

README.md

@@ -148,7 +148,7 @@ Master-skill 是由 [FoJin](https://fojin.app) 驱动的佛教 AI 祖师人格
 - **渐进式披露**：SKILL.md 以决策树 + Quick Ref 为主，`references/`、`sources/` 按需加载，Context 随查随取
 - **HARD-GATE 铁律**：`/create-master` 与预置法师内置红线——教义断言、修行指导与文本解释必须引用该 persona 声明的来源（CBETA / BDRC / Toh / SuttaCentral / PTS / 合规编纂开示），不得捏造来源 ID，不得为虚构人物建角色
 - **二阶段独立审查**：生成管线在写入前强制经过"教义准确性 → 风格一致性"两轮独立审查，FAIL 自动修复最多 2 轮
-- **自动化保真度测试**：每位祖师 `tests/fidelity.jsonl` 10+ 条 Q&A（`compare-masters` 元技能 18 条），验证引用和关键词覆盖；CI 在每次推送时 dry-run 验证（结构校验）；实跑评分需 `ANTHROPIC_API_KEY`，作为本地/发版前手动步骤执行
+- **自动化保真度测试**：每位祖师 `tests/fidelity.jsonl` 10+ 条 Q&A（`compare-masters` 元技能 18 条），验证引用和关键词覆盖；CI 在每次推送时 dry-run 验证（结构校验）；实跑评分需 `ANTHROPIC_API_KEY`，作为本地/发版前手动步骤执行——首份[实测基线](#保真度基线首次实测)已提交：59/84 已测通过（70%），全量 211 条覆盖率 40%（详见 [eval/reports/](eval/reports/)）
 - **多平台统一插件**：Claude Code、Cursor、Codex CLI、OpenCode、Gemini CLI 共用一份 `prebuilt/`，session-start hook 跨平台注入法师列表
 - **NPX 一键安装**：`npx master-skill install master-zhiyi` 直接部署到 Claude Code
 - **离线工具链**：`scripts/cite.py`（CBETA 引用查询）、`scripts/query.py`（离线语义检索）、`scripts/validate.py`（frontmatter linter）
@@ -163,11 +163,24 @@ Master-skill 的核心不是"角色扮演提示词集合"，而是一个可验
 |---|---|
 | 有来源 | 每位祖师声明 `sources[]`、离线 excerpts、FoJin live fallback 与引用自审 |
 | 守边界 | `ETHICS.md`、每位祖师 Layer 0 HARD-GATE、版权 Tier 与教界越界报告机制 |
-| 可评测 | `tests/fidelity.jsonl`、persona-fidelity schema、promptfoo RAW / SPE / CUS 评测层 |
+| 可评测 | `tests/fidelity.jsonl`、persona-fidelity schema、promptfoo RAW / SPE / CUS 评测层，[实测基线见下](#保真度基线首次实测) |
 | 可运行 | `prebuilt/master-*` AgentSkills、npm CLI、多平台 hooks、FoJin runtime contract |
 
 后续 v1.0 路线以框架稳定为优先：见 [docs/v1-framework-roadmap.md](docs/v1-framework-roadmap.md) 与 [docs/fojin-runtime-contract.md](docs/fojin-runtime-contract.md)。
 
+### 保真度基线（首次实测）
+
+`tests/fidelity.jsonl` 曾经只是"存在的夹具"——`scripts/test-fidelity.py` 只打印到终端，仓库里从未提交过一次真实评分。2026-08-18 首次跑出并提交了这份基线（commit [`c697d5d`](https://github.com/xr843/Master-skill/commit/c697d5d3be78ce6738cf1f969ca057c7e4c16bb5)，模型 `claude-sonnet-4-6`）：
+
+| | 数值 |
+|---|---|
+| 已测通过 / 已测总数 | **59 / 84（70%）** |
+| 全量夹具覆盖率 | 84 / 211（40%）—— 运行途中 API 账户余额耗尽（HTTP 400），非限流也非代码缺陷，剩余 127 条**未测**，不计入失败 |
+| 真实失败聚集 | 关键词未覆盖 14/25；禁用词命中 12/25；引用缺失 5/25；**零虚构引用** |
+| ⚠️ 量具告警 | 12 条禁用词失败里有 **10 条，禁用词本就出现在提问中**（陷阱题），而检查是对回答做纯子串匹配——正确的驳斥与真正的越界会被同等判失败。**故 70% 是下限而非估计值**，详见基线报告 |
+
+这是**关键词/引用字符串覆盖率检查，不是教义正确性或 LLM 判分的答案质量**。完整表格、失败案例与方法论说明见 **[eval/reports/BASELINE.md](eval/reports/BASELINE.md)**。
+
 ---
 
 ## 开发者安装
@@ -341,9 +354,9 @@ git clone https://github.com/xr843/Master-skill ~/Master-skill
 
 ## 桌面管理器
 
-原生桌面控制台(纯 Rust,egui,单二进制,无 Electron),统一管理 17 个 master skill 的安装状态、fidelity 评测覆盖率、运行追踪与质量门禁:
+原生桌面控制台(纯 Rust,egui,单二进制,无 Electron),统一管理 19 个 master skill 的安装状态、fidelity 评测覆盖率、运行追踪与质量门禁:
 
-![Master-skill Desktop Manager](https://raw.githubusercontent.com/xr843/Master-skill/master/docs/assets/desktop-manager.png)
+![Master-skill Desktop Manager](https://raw.githubusercontent.com/xr843/Master-skill/main/docs/assets/desktop-manager.png)
 
 **下载**:[Releases](https://github.com/xr843/Master-skill/releases) 提供 Linux / Windows / macOS 预编译二进制,下载后直接运行(仓库根目录下执行,需本地已 clone 本仓库)。Linux / macOS 下载后需先 `chmod +x` 赋予可执行权限;macOS 上二进制未签名,首次运行需右键"打开"或执行 `xattr -d com.apple.quarantine <文件名>` 解除隔离。
 

README_EN.md

@@ -146,7 +146,7 @@ This project is built out of respect for Buddhist traditions. All content is gen
 - **Progressive disclosure**: SKILL.md is a decision tree + quick reference; `references/` and `sources/` are loaded on demand to keep context lean
 - **HARD-GATE discipline**: Both `/create-master` and every prebuilt master require doctrinal claims, practice guidance, and text interpretation to cite that persona's declared sources (CBETA / BDRC / Toh / SuttaCentral / PTS / compliant compiled teachings); fabricated source IDs and fictional personas are forbidden
 - **Two-stage independent review**: The generation pipeline forces a "doctrinal accuracy → voice consistency" review before write; FAIL triggers up to 2 rounds of automatic repair
-- **Automated fidelity tests**: Each master's `tests/fidelity.jsonl` holds 10+ Q&A samples (the `compare-masters` meta-skill holds 18) validating citations and keyword coverage; CI runs a dry-run on every push
+- **Automated fidelity tests**: Each master's `tests/fidelity.jsonl` holds 10+ Q&A samples (the `compare-masters` meta-skill holds 18) validating citations and keyword coverage; CI runs a dry-run on every push, and a graded run needs `ANTHROPIC_API_KEY` as a manual local/pre-release step — the first committed [baseline](#fidelity-baseline-first-real-run) landed 2026-08-18: 59/84 measured cases passed (70%), covering 40% of all 211 fixtures (see [eval/reports/](eval/reports/))
 - **Unified multi-platform plugin**: Claude Code, Cursor, Codex CLI, OpenCode, and Gemini CLI share one `prebuilt/` tree, with a session-start hook injecting the master list on every platform
 - **NPX one-shot install**: `npx master-skill install master-zhiyi` drops skills straight into Claude Code
 - **Offline toolchain**: `scripts/cite.py` (CBETA lookup), `scripts/query.py` (offline semantic search), `scripts/validate.py` (frontmatter linter)
@@ -161,11 +161,24 @@ Master-skill is not a prompt pack. It is a verifiable Buddhist AI persona framew
 |---|---|
 | Source-grounded | `sources[]`, offline excerpts, FoJin live fallback, and citation self-audits per master |
 | Boundary-aware | `ETHICS.md`, per-master Layer 0 HARD-GATE rules, copyright tiers, and boundary violation reporting |
-| Fidelity-tested | `tests/fidelity.jsonl`, persona-fidelity schema, and promptfoo RAW / SPE / CUS evals |
+| Fidelity-tested | `tests/fidelity.jsonl`, persona-fidelity schema, promptfoo RAW / SPE / CUS evals, [real measured baseline below](#fidelity-baseline-first-real-run) |
 | Runtime-ready | `prebuilt/master-*` AgentSkills, npm CLI, multi-platform hooks, and a FoJin runtime contract |
 
 The v1.0 track prioritizes framework stability over adding more masters. See [docs/v1-framework-roadmap.md](docs/v1-framework-roadmap.md) and [docs/fojin-runtime-contract.md](docs/fojin-runtime-contract.md).
 
+### Fidelity baseline (first real run)
+
+The 211 fixtures under `tests/fidelity.jsonl` used to be just fixtures — `scripts/test-fidelity.py` only printed to stdout, and no scored run had ever been committed. On 2026-08-18 we ran and committed the first real baseline (commit [`c697d5d`](https://github.com/xr843/Master-skill/commit/c697d5d3be78ce6738cf1f969ca057c7e4c16bb5), model `claude-sonnet-4-6`):
+
+| | Value |
+|---|---|
+| Passed / measured | **59 / 84 (70%)** |
+| Coverage of the full suite | 84 / 211 (40%) — the run stopped partway when the API account's credit balance ran out (HTTP 400), not from rate limiting or a code bug; the remaining 127 cases were never evaluated and are **not** counted as failures |
+| Where real failures cluster | Missing expected keyword 14/25; forbidden phrase present 12/25; missing citation 5/25; **zero fabricated citations** |
+| ⚠️ Instrument warning | 10 of the 12 forbidden-phrase failures name a term **that already appears in the question** (these are trap questions), and the check is a plain substring match on the response — a correct refusal fails it exactly as hard as a real violation. **So 70% is a floor, not an estimate.** See the baseline report |
+
+This measures **keyword/citation-string coverage, not doctrinal correctness or LLM-judged answer quality**. Full table, failing cases, and methodology notes: **[eval/reports/BASELINE.md](eval/reports/BASELINE.md)**.
+
 ---
 
 ## Developer Installation
@@ -328,9 +341,9 @@ The system will guide you through a three-step intake, then automatically collec
 
 ## Desktop Manager
 
-A native desktop console (pure Rust, egui, single binary, no Electron) that unifies management of installation status, fidelity evaluation coverage, run tracing, and the quality gate across all 17 master skills:
+A native desktop console (pure Rust, egui, single binary, no Electron) that unifies management of installation status, fidelity evaluation coverage, run tracing, and the quality gate across all 19 master skills:
 
-![Master-skill Desktop Manager](https://raw.githubusercontent.com/xr843/Master-skill/master/docs/assets/desktop-manager.png)
+![Master-skill Desktop Manager](https://raw.githubusercontent.com/xr843/Master-skill/main/docs/assets/desktop-manager.png)
 
 **Download**: [Releases](https://github.com/xr843/Master-skill/releases) provides pre-built binaries for Linux / Windows / macOS — download and run directly (execute from the repository root; requires a local clone of this repo). On Linux/macOS you'll need to `chmod +x` the downloaded binary first; on macOS it's unsigned, so the first run needs right-click → Open, or `xattr -d com.apple.quarantine <file>` to clear the quarantine flag.