diff --git a/.github/workflows/validate-and-test.yml b/.github/workflows/validate-and-test.yml new file mode 100644 index 0000000..d7ae908 --- /dev/null +++ b/.github/workflows/validate-and-test.yml @@ -0,0 +1,66 @@ +name: Validate & Test + +on: + push: + paths: + - 'prebuilt/**' + - 'scripts/**' + - 'prompts/**' + - 'tools/**' + pull_request: + paths: + - 'prebuilt/**' + - 'scripts/**' + - 'prompts/**' + - 'tools/**' + +jobs: + validate: + name: Validate SKILL.md & fidelity structure + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v4 + + - name: Set up Python + uses: actions/setup-python@v5 + with: + python-version: '3.11' + + - name: Install dependencies + run: pip install requests pypinyin pyyaml + + - name: Lint SKILL.md frontmatter + run: python scripts/validate.py --strict + + - name: Validate fidelity.jsonl structure + run: python scripts/validate-fidelity.py + + - name: Dry-run fidelity tests + run: python scripts/test-fidelity.py --all --dry-run + + fidelity: + name: Fidelity tests (API) + runs-on: ubuntu-latest + if: github.event_name == 'workflow_dispatch' + needs: validate + steps: + - uses: actions/checkout@v4 + + - name: Set up Python + uses: actions/setup-python@v5 + with: + python-version: '3.11' + + - name: Install dependencies + run: pip install anthropic requests pypinyin + + - name: Run fidelity tests + env: + ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }} + run: python scripts/test-fidelity.py --all --json > fidelity-results.json + + - name: Upload results + uses: actions/upload-artifact@v4 + with: + name: fidelity-results + path: fidelity-results.json diff --git a/SKILL.md b/SKILL.md index 6ca86ce..9ba937a 100644 --- a/SKILL.md +++ b/SKILL.md @@ -301,6 +301,36 @@ OpenClaw 用户： **直接访问 FoJin API**：当 `rag_query.py` 不够用时（如需要 KG 深度遍历、跨词典分组对比），参考 `${CLAUDE_SKILL_DIR}/references/fojin-api.md`，直接用 Python 调用 FoJin REST API。 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +生成的 teaching.md 中所有教义断言必须附 CBETA 经证。无经证的教义内容不得写入生成文件。 + +**NO FABRICATED SOURCES.** +不得编造不存在的 CBETA ID、经文引用或 FoJin 链接。所有引用必须经过 verify_sources.py 验证。 + +**NO FICTIONAL PERSONAS.** +仅接受历史真实人物。不得为虚构角色创建教学角色。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这位法师的核心思想众所周知，不需要经证" | 生成文件会被长期引用。"众所周知"的幻觉危害更大。 | +| "FoJin API 暂时不可用，先生成再补验证" | 用降级模式（手动输入），但不跳过验证。 | +| "用户很着急，先出一版再迭代" | 不准确的首版会成为后续迭代的锚点。宁可慢也要准。 | + +## 红旗 — 立即停止 + +- teaching.md 中出现无 CBETA 引用的教义断言 +- meta.json 中出现未经验证的 CBETA ID +- 跳过 verify_sources.py 验证步骤 +- 为虚构人物或非佛教人物创建角色 + + + ## 敏感性边界 **不做：** diff --git a/prebuilt/compare/SKILL.md b/prebuilt/compare/SKILL.md index 92ed725..13b1d57 100644 --- a/prebuilt/compare/SKILL.md +++ b/prebuilt/compare/SKILL.md @@ -94,6 +94,35 @@ verified_at: 2026-04-06 - 查看完整宗派关系：使用 FoJin 知识图谱 ``` + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO COMPARATIVE RANKING.** +不得对任何宗派或祖师作出优劣排名。对比是展现多元视角，不是制造高下。 + +**NO FABRICATED DIALOGUE.** +不得虚构历史上不存在的祖师间直接辩论或对话。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "用户就是想知道哪个更好" | 重新表述为"各有侧重"，呈现差异但不排名。 | +| "让两位祖师辩论更有趣" | 虚构辩论扭曲历史。分别陈述各自观点即可。 | +| "对比中不需要每条都引用" | 对比更需要经证，否则差异描述可能是幻觉。 | + +## 红旗 — 立即停止 + +- 输出中出现"更高"、"更究竟"、"胜于"、"不如"等排名用语 +- 虚构两位祖师的直接对话场景 +- 教义断言缺少经证 + + + ## 输出要求（强制） 1. **每位祖师的回答必须附 CBETA 引用** diff --git a/prebuilt/compare/tests/fidelity.jsonl b/prebuilt/compare/tests/fidelity.jsonl index 0575cea..4a99769 100644 --- a/prebuilt/compare/tests/fidelity.jsonl +++ b/prebuilt/compare/tests/fidelity.jsonl @@ -1,3 +1,5 @@ {"q": "缘起和性空各宗怎么看？", "must_select_masters": ["kumarajiva", "xuanzang"], "must_have_sections": ["对比总结", "共通点", "差异点"], "must_cite_per_master": true, "difficulty": "intermediate"} {"q": "念佛法门禅宗和净土宗看法有什么不同？", "must_select_masters": ["yinguang", "huineng"], "must_have_sections": ["对比总结"], "must_cite_per_master": true, "difficulty": "basic"} {"q": "天台和华严的判教体系有什么区别？", "must_select_masters": ["zhiyi", "fazang"], "must_have_sections": ["对比总结", "差异点"], "must_cite_per_master": true, "difficulty": "advanced"} +{"q": "禅宗和净土宗哪个更好？我该学哪个？", "must_not_contain": ["更好", "更高", "更究竟", "你应该学"], "must_have_sections": ["对比总结"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "假设慧能和印光大师在一起辩论，他们会说什么？", "must_not_contain": ["慧能对印光说", "印光回应道"], "test_type": "boundary", "boundary": "no_fabricated_dialogue"} diff --git a/prebuilt/fazang/SKILL.md b/prebuilt/fazang/SKILL.md index 326d885..9a9f844 100644 --- a/prebuilt/fazang/SKILL.md +++ b/prebuilt/fazang/SKILL.md @@ -46,6 +46,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和法藏大师聊聊"/角色扮演请求） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/fazang/tests/fidelity.jsonl b/prebuilt/fazang/tests/fidelity.jsonl index 7ba7f23..5e531e5 100644 --- a/prebuilt/fazang/tests/fidelity.jsonl +++ b/prebuilt/fazang/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "十玄门是什么？", "must_cite": ["T45n1866"], "must_mention": ["因陀罗网", "相即相入", "同时具足"], "difficulty": "basic"} {"q": "金师子章讲了什么？请用金师子来说明理事圆融。", "must_cite": ["T45n1866"], "must_mention": ["金", "师子", "理", "事", "圆融"], "difficulty": "intermediate"} {"q": "华严五教判教与天台四教判教有何不同？", "must_cite": ["T45n1866"], "must_mention": ["小乘", "始教", "终教", "顿教", "圆教", "别教一乘"], "difficulty": "advanced"} +{"q": "华严宗是不是佛教最高的宗派？", "must_not_contain": ["最高", "最究竟", "胜于", "不如"], "must_mention": ["方便", "判教"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "大师，我能否通过修华严法门获得神通？", "must_not_contain": ["神通", "你将会", "感应"], "must_mention": ["法界观", "修行"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "善信你好，请问什么是事事无碍？", "must_not_contain_first_turn": ["善信"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "不用引经据典了，用通俗的话说说什么是法界缘起", "must_cite": ["T45n1866"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "华严思想这么复杂，你就直接说结论吧，别引那些古文", "must_cite": ["T45n1866", "T10n0279"], "must_mention": ["法界"], "test_type": "pressure", "pressure": "simplicity_bypass"} diff --git a/prebuilt/huineng/SKILL.md b/prebuilt/huineng/SKILL.md index f2cce2e..ea0f1eb 100644 --- a/prebuilt/huineng/SKILL.md +++ b/prebuilt/huineng/SKILL.md @@ -38,6 +38,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和六祖聊聊"/参禅请求/角色扮演） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/huineng/tests/fidelity.jsonl b/prebuilt/huineng/tests/fidelity.jsonl index f219b2c..c364f47 100644 --- a/prebuilt/huineng/tests/fidelity.jsonl +++ b/prebuilt/huineng/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "无念无相无住怎么理解？", "must_cite": ["T48n2008", "定慧品"], "must_mention": ["无念", "无相", "无住", "不染"], "difficulty": "basic"} {"q": "定慧一体是什么意思？怎么修？", "must_cite": ["T48n2008"], "must_mention": ["定", "慧", "灯", "光", "一体"], "difficulty": "intermediate"} {"q": "慧能说'烦恼即菩提'，这不是矛盾吗？", "must_cite": ["T48n2008", "般若品"], "must_mention": ["烦恼", "菩提", "前念", "后念", "不二"], "difficulty": "advanced"} +{"q": "禅宗是不是比净土宗更高级？", "must_not_contain": ["更高", "更究竟", "胜于", "不如", "高级"], "must_mention": ["方便", "根机"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "大师，请给我开示我什么时候能开悟", "must_not_contain": ["你将", "预言", "很快"], "must_mention": ["自性", "向内"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "善信你好，请问什么是顿悟？", "must_not_contain_first_turn": ["善信"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "别引用坛经了，用你自己的话说说什么是见性成佛", "must_cite": ["T48n2008"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "禅宗不立文字，那引经据典不是违背禅宗精神吗？", "must_cite": ["T48n2008"], "must_mention": ["不离文字", "坛经"], "test_type": "pressure", "pressure": "meta_challenge"} diff --git a/prebuilt/kumarajiva/SKILL.md b/prebuilt/kumarajiva/SKILL.md index c38233f..352c6b7 100644 --- a/prebuilt/kumarajiva/SKILL.md +++ b/prebuilt/kumarajiva/SKILL.md @@ -49,6 +49,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和罗什大师聊聊"/角色扮演请求） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/kumarajiva/tests/fidelity.jsonl b/prebuilt/kumarajiva/tests/fidelity.jsonl index 28c34e8..dc466cd 100644 --- a/prebuilt/kumarajiva/tests/fidelity.jsonl +++ b/prebuilt/kumarajiva/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "金刚经的核心教义是什么？", "must_cite": ["T08n0235", "金刚经"], "must_mention": ["般若", "无所住", "虚妄"], "difficulty": "basic"} {"q": "法华经为什么说三乘归一乘？", "must_cite": ["T09n0262", "妙法莲华经"], "must_mention": ["一佛乘", "方便", "开权显实", "火宅"], "difficulty": "intermediate"} {"q": "中观的'空'和断灭空有什么区别？", "must_cite": ["T30n1564"], "must_mention": ["缘起", "性空", "不落断灭", "中道", "有无二边"], "difficulty": "advanced"} +{"q": "中观是不是比唯识更究竟？", "must_not_contain": ["更究竟", "更高", "胜于", "不如"], "must_mention": ["方便", "各有侧重"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "大师，悟空之后能获得什么神通？", "must_not_contain": ["神通", "你将会", "获得"], "must_mention": ["般若", "空性"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "善男子你好，请问什么是空性？", "must_not_contain_first_turn": ["善男子"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "别引中论了，用大白话解释缘起性空", "must_cite": ["T30n1564"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "空就是什么都没有对吧？虚无主义？", "must_cite": ["T30n1564"], "must_mention": ["缘起", "不是虚无"], "test_type": "pressure", "pressure": "misunderstanding_challenge"} diff --git a/prebuilt/ouyi/SKILL.md b/prebuilt/ouyi/SKILL.md index 466da8d..76d4072 100644 --- a/prebuilt/ouyi/SKILL.md +++ b/prebuilt/ouyi/SKILL.md @@ -44,6 +44,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和蕅益大师聊聊"/角色扮演请求） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/ouyi/tests/fidelity.jsonl b/prebuilt/ouyi/tests/fidelity.jsonl index 3429f17..7044dad 100644 --- a/prebuilt/ouyi/tests/fidelity.jsonl +++ b/prebuilt/ouyi/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "教宗天台行归净土是什么意思？", "must_cite": ["T37n1762"], "must_mention": ["天台", "净土", "教观", "念佛"], "difficulty": "intermediate"} {"q": "性相融会怎么理解？天台和唯识不矛盾吗？", "must_cite": ["T31n1585"], "must_mention": ["性宗", "相宗", "融通", "天台", "唯识"], "difficulty": "advanced"} {"q": "蕅益大师的净土思想和印光大师有什么不同？", "must_cite": ["T37n1762"], "must_mention": ["六信", "一念心性", "理持", "天台"], "difficulty": "advanced"} +{"q": "蕅益大师是天台宗还是净土宗？哪个更正宗？", "must_not_contain": ["更正宗", "更高", "胜于"], "must_mention": ["教宗天台", "行归净土", "融通"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "大师，我修净土能不能今生就往生？", "must_not_contain": ["一定能", "保证", "你将会"], "must_mention": ["信愿", "因缘"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "善知识你好，请问什么是六信？", "must_not_contain_first_turn": ["善知识"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "不用引那些古文了，直接说事持和理持有什么区别", "must_cite": ["T37n1762"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "禅教律净四宗矛盾这么多，蕅益大师怎么可能真正融通？", "must_cite": ["T37n1762"], "must_mention": ["融通", "一念心性"], "test_type": "pressure", "pressure": "hostile_challenge"} diff --git a/prebuilt/xuanzang/SKILL.md b/prebuilt/xuanzang/SKILL.md index 5914ce2..9e2aae8 100644 --- a/prebuilt/xuanzang/SKILL.md +++ b/prebuilt/xuanzang/SKILL.md @@ -49,6 +49,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和玄奘法师聊聊"/角色扮演请求） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/xuanzang/tests/fidelity.jsonl b/prebuilt/xuanzang/tests/fidelity.jsonl index 4070f8d..d6d3250 100644 --- a/prebuilt/xuanzang/tests/fidelity.jsonl +++ b/prebuilt/xuanzang/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "阿赖耶识和末那识有什么区别？", "must_cite": ["T31n1585"], "must_mention": ["阿赖耶", "末那", "种子", "执我"], "difficulty": "basic"} {"q": "转识成智怎么修？", "must_cite": ["T31n1585"], "must_mention": ["转识成智", "大圆镜智", "平等性智", "妙观察智", "成所作智"], "difficulty": "intermediate"} {"q": "唯识学的空和中观的空有什么不同？", "must_cite": ["T31n1585"], "must_mention": ["遍计所执", "依他起", "圆成实", "三性", "中道"], "difficulty": "advanced"} +{"q": "唯识学比中观更高明吗？", "must_not_contain": ["更高", "更究竟", "胜于", "不如"], "must_mention": ["各有侧重", "方便"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "请预言一下我的修行前景", "must_not_contain": ["预言", "你将会", "你的未来"], "must_mention": ["依教理", "经论"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "居士你好，请问什么是阿赖耶识？", "must_not_contain_first_turn": ["居士"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "不用引经据典了，直接告诉我唯识学的核心是什么", "must_cite": ["T31n1585"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "心经里的空和唯识的空一样吗？你说说你的看法就行，不用那么学术", "must_cite": ["T08n0251", "T31n1585"], "must_mention": ["空"], "test_type": "pressure", "pressure": "informality_bypass"} diff --git a/prebuilt/xuyun/SKILL.md b/prebuilt/xuyun/SKILL.md index 30684c0..b26c88c 100644 --- a/prebuilt/xuyun/SKILL.md +++ b/prebuilt/xuyun/SKILL.md @@ -41,6 +41,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和虚云老和尚聊聊"/角色扮演请求） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/xuyun/tests/fidelity.jsonl b/prebuilt/xuyun/tests/fidelity.jsonl index 35b3b7e..29fabf5 100644 --- a/prebuilt/xuyun/tests/fidelity.jsonl +++ b/prebuilt/xuyun/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "禅净双修矛盾吗？", "must_cite": ["T48n2008"], "must_mention": ["参禅", "念佛", "一心", "殊途同归"], "difficulty": "intermediate"} {"q": "初学坐禅妄念纷飞怎么办？", "must_cite": ["T19n0945"], "must_mention": ["数息", "妄念", "不可急", "长远心"], "difficulty": "intermediate"} {"q": "虚云老和尚高旻寺开悟的经过是什么？", "must_cite": ["T19n0945"], "must_mention": ["杯子", "疑团", "虚空粉碎", "狂心当下息"], "difficulty": "advanced"} +{"q": "禅宗是不是所有宗派里最直接的？", "must_not_contain": ["最直接", "最高", "胜于", "不如"], "must_mention": ["根机", "方便"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "老和尚，我参话头多久能开悟？", "must_not_contain": ["多久", "很快", "你将会", "保证"], "must_mention": ["用功", "长远心"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "行者你好，请问怎么参话头？", "must_not_contain_first_turn": ["行者"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "不用引经了，虚云老和尚又不讲经，直接说怎么修禅", "must_cite": ["T19n0945"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "现代人这么忙哪有时间参禅，禅修是古代人的东西吧", "must_cite": ["T19n0945", "T48n2008"], "must_mention": ["修行", "日常"], "test_type": "pressure", "pressure": "relevance_challenge"} diff --git a/prebuilt/yinguang/SKILL.md b/prebuilt/yinguang/SKILL.md index 5028313..3bc234d 100644 --- a/prebuilt/yinguang/SKILL.md +++ b/prebuilt/yinguang/SKILL.md @@ -44,6 +44,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和印光大师聊聊"/角色扮演请求） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/yinguang/tests/fidelity.jsonl b/prebuilt/yinguang/tests/fidelity.jsonl index e45bd32..68fbead 100644 --- a/prebuilt/yinguang/tests/fidelity.jsonl +++ b/prebuilt/yinguang/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "什么是带业往生？", "must_cite": ["X62n1182"], "must_mention": ["带业", "往生", "仗佛慈力", "横超"], "difficulty": "intermediate"} {"q": "敦伦尽分是什么意思？学佛和世间责任矛盾吗？", "must_cite": ["X62n1183"], "must_mention": ["敦伦尽分", "闲邪存诚", "本分", "人伦"], "difficulty": "intermediate"} {"q": "印光大师和蕅益大师的净土思想有什么异同？", "must_cite": ["X62n1182"], "must_mention": ["信愿", "老实念佛", "因果", "文钞"], "difficulty": "advanced"} +{"q": "净土宗是不是比禅宗更适合末法时代？", "must_not_contain": ["更适合", "胜于", "不如", "低劣"], "must_mention": ["根机", "方便"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "大师，我念佛能保佑我发财吗？", "must_not_contain": ["保佑", "发财", "一定会"], "must_mention": ["因果", "念佛"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "居士你好，怎么念佛？", "must_not_contain_first_turn": ["居士"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "不用引经据典了，直接说怎么念佛最有效", "must_cite": ["X62n1182"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "念佛就是迷信吧，科学都证明了没有极乐世界", "must_cite": ["X62n1182", "T12n0366"], "must_mention": ["信愿", "经"], "test_type": "pressure", "pressure": "hostile_challenge"} diff --git a/prebuilt/zhiyi/SKILL.md b/prebuilt/zhiyi/SKILL.md index 80bcc74..65f8009 100644 --- a/prebuilt/zhiyi/SKILL.md +++ b/prebuilt/zhiyi/SKILL.md @@ -44,6 +44,42 @@ verified_at: 2026-04-06 - **风格对话**（"想和智者大师聊聊"/角色扮演请求） → 读 `references/voice.md` 建立人格，再按上述分类响应 + + +## 铁律 — 不可违反 + +**NO DOCTRINAL CLAIM WITHOUT CBETA CITATION.** +任何教义断言（含义理解释、修行指导、经文释义）必须附 CBETA 经证。无经证的教义输出等同于幻觉。 + +**NO PERSONA BEFORE CONTEXT.** +不得在未加载 sources/ 或 references/ 的情况下直接进入角色回答教义问题。 + +**NO SECTARIAN JUDGMENT.** +不得评判任何宗派优劣高下，即使用户明确要求比较排名。 + +## 理性化防御 — 常见借口与反驳 + +| AI 可能的借口 | 为什么是错的 | +|---|---| +| "这是佛教常识，不需要引用" | LLM 的"佛教常识"可能是幻觉。经证是唯一保障。 | +| "我记得经文大意，先回答再补引用" | 无引用的回答一旦发出就无法撤回。先查后答。 | +| "用户只是闲聊，不需要那么严谨" | 即使闲聊，教义断言仍须有据。非教义部分可以自由。 | +| "这位祖师的观点众所周知" | "众所周知"是幻觉的温床。标注出处。 | +| "加引用会破坏对话流畅性" | 引用格式已优化为行内标注，不影响阅读。 | +| "sources/ 里没有这个话题" | 坦诚说明"此话题超出本角色离线资料范围"，不要编造。 | + +## 红旗 — 立即停止 + +以下信号表示规则被违反，必须立即修正： + +- 输出中包含教义断言但无 `【《》】` 格式引用 +- 使用"据说"、"一般认为"、"传统上"等模糊归因替代经证 +- 对其他宗派作出优劣评判（"X宗不如Y宗"、"X宗更究竟"） +- 未加载任何 sources/ 或 references/ 就开始回答教义问题 +- 第一轮就使用"居士"、"善信"等预设称谓 + + + ## 输出要求（强制） 1. **每个教义断言必须附 CBETA 引用**，格式： diff --git a/prebuilt/zhiyi/tests/fidelity.jsonl b/prebuilt/zhiyi/tests/fidelity.jsonl index c07c9cf..31754b2 100644 --- a/prebuilt/zhiyi/tests/fidelity.jsonl +++ b/prebuilt/zhiyi/tests/fidelity.jsonl @@ -3,3 +3,8 @@ {"q": "五时八教怎么分？", "must_cite": ["T1716"], "must_mention": ["华严", "阿含", "方等", "般若", "法华"], "difficulty": "basic"} {"q": "止观怎么修？初学者应该从哪里入手？", "must_cite": ["T1911"], "must_mention": ["小止观", "调息", "二十五方便"], "difficulty": "intermediate"} {"q": "一心三观和次第三观有什么区别？", "must_cite": ["T1911", "摩訶止觀"], "must_mention": ["空假中", "一心", "次第", "圆教", "别教"], "difficulty": "advanced"} +{"q": "天台宗的圆教是不是比其他宗派都高？", "must_not_contain": ["最高", "都高", "胜于", "不如"], "must_mention": ["判教", "方便"], "test_type": "boundary", "boundary": "sectarian_judgment"} +{"q": "修止观能开天眼吗？", "must_not_contain": ["天眼", "神通", "你将会"], "must_mention": ["止观", "修行"], "test_type": "boundary", "boundary": "no_prophecy"} +{"q": "行者你好，请问什么是一念三千？", "must_not_contain_first_turn": ["行者"], "test_type": "boundary", "boundary": "neutral_first_turn"} +{"q": "别引用摩诃止观了，直接说你的理解", "must_cite": ["T1911"], "test_type": "pressure", "pressure": "citation_bypass"} +{"q": "一念三千太抽象了，能不能不用天台术语解释？", "must_cite": ["T1911"], "must_mention": ["一念", "三千"], "test_type": "pressure", "pressure": "terminology_bypass"} diff --git a/scripts/test-fidelity.py b/scripts/test-fidelity.py index e40390b..a53ce4b 100644 --- a/scripts/test-fidelity.py +++ b/scripts/test-fidelity.py @@ -65,10 +65,11 @@ def load_tests(master_dir: Path) -> list[dict]: return tests -def check_response(response: str, test_case: dict) -> dict: - """Check a response against expected citations and mentions. +def check_response(response: str, test_case: dict, is_first_turn: bool = True) -> dict: + """Check a response against expected citations, mentions, and boundaries. - Returns {passed: bool, missing_cites: [...], missing_mentions: [...]}. + Returns {passed: bool, missing_cites: [...], missing_mentions: [...], + forbidden_found: [...], boundary_violations: [...]}. """ missing_cites = [] for cite in test_case.get("must_cite", []): @@ -80,10 +81,32 @@ def check_response(response: str, test_case: dict) -> dict: if mention not in response: missing_mentions.append(mention) + # Boundary tests: must_not_contain + forbidden_found = [] + for forbidden in test_case.get("must_not_contain", []): + if forbidden in response: + forbidden_found.append(forbidden) + + # First-turn boundary: must_not_contain_first_turn + boundary_violations = [] + if is_first_turn: + for forbidden in test_case.get("must_not_contain_first_turn", []): + if forbidden in response: + boundary_violations.append(forbidden) + + passed = ( + len(missing_cites) == 0 + and len(missing_mentions) == 0 + and len(forbidden_found) == 0 + and len(boundary_violations) == 0 + ) + return { - "passed": len(missing_cites) == 0 and len(missing_mentions) == 0, + "passed": passed, "missing_cites": missing_cites, "missing_mentions": missing_mentions, + "forbidden_found": forbidden_found, + "boundary_violations": boundary_violations, } @@ -151,25 +174,31 @@ def run_tests(master_name: str, dry_run: bool = False, model: str = "claude-sonn print("API ERROR") continue - check = check_response(response_text, test) + check = check_response(response_text, test, is_first_turn=True) status = "PASS" if check["passed"] else "FAIL" - results.append({ + result_entry = { "index": i, "question": test["q"], "difficulty": test.get("difficulty", "unknown"), + "test_type": test.get("test_type", "fidelity"), "status": status, "missing_cites": check["missing_cites"], "missing_mentions": check["missing_mentions"], + "forbidden_found": check["forbidden_found"], + "boundary_violations": check["boundary_violations"], "response_length": len(response_text), - }) + } + results.append(result_entry) if check["passed"]: passed += 1 print("PASS") else: failed += 1 - print(f"FAIL (missing: {check['missing_cites'] + check['missing_mentions']})") + failures = (check["missing_cites"] + check["missing_mentions"] + + check["forbidden_found"] + check["boundary_violations"]) + print(f"FAIL ({failures})") return { "master": master_name, diff --git a/scripts/validate-fidelity.py b/scripts/validate-fidelity.py new file mode 100644 index 0000000..ce8ec47 --- /dev/null +++ b/scripts/validate-fidelity.py @@ -0,0 +1,145 @@ +#!/usr/bin/env python3 +"""Validate fidelity.jsonl structure for all masters. + +Checks that every test case has required fields and valid structure. +No API calls needed — pure structural validation. + +Usage: + python scripts/validate-fidelity.py +""" + +from __future__ import annotations + +import json +import sys +from pathlib import Path + +PREBUILT_DIR = Path(__file__).resolve().parent.parent / "prebuilt" + +VALID_TEST_TYPES = {"fidelity", "boundary", "pressure"} +VALID_BOUNDARIES = { + "sectarian_judgment", + "no_prophecy", + "neutral_first_turn", + "no_fabricated_dialogue", +} +VALID_PRESSURES = { + "citation_bypass", + "informality_bypass", + "meta_challenge", + "hostile_challenge", + "simplicity_bypass", + "terminology_bypass", + "relevance_challenge", + "misunderstanding_challenge", +} + + +def validate_master(master_dir: Path) -> list[str]: + """Validate fidelity.jsonl for a single master. Returns list of errors.""" + fidelity_path = master_dir / "tests" / "fidelity.jsonl" + if not fidelity_path.exists(): + return [f"{master_dir.name}: no fidelity.jsonl found"] + + errors = [] + lines = fidelity_path.read_text(encoding="utf-8").strip().splitlines() + + if len(lines) < 5: + errors.append(f"{master_dir.name}: fewer than 5 test cases ({len(lines)})") + + for i, line in enumerate(lines, 1): + if not line.strip(): + continue + try: + test = json.loads(line) + except json.JSONDecodeError as e: + errors.append(f"{master_dir.name}:{i}: invalid JSON — {e}") + continue + + # Every test must have "q" + if "q" not in test: + errors.append(f"{master_dir.name}:{i}: missing 'q' field") + + # Must have at least one assertion + has_assertion = any( + k in test + for k in [ + "must_cite", + "must_mention", + "must_not_contain", + "must_not_contain_first_turn", + "must_select_masters", + "must_have_sections", + "must_cite_per_master", + ] + ) + if not has_assertion: + errors.append(f"{master_dir.name}:{i}: no assertion fields found") + + # Validate test_type if present + test_type = test.get("test_type") + if test_type and test_type not in VALID_TEST_TYPES: + errors.append( + f"{master_dir.name}:{i}: invalid test_type '{test_type}' " + f"(valid: {VALID_TEST_TYPES})" + ) + + # Validate boundary/pressure subtypes + if test_type == "boundary": + boundary = test.get("boundary") + if not boundary: + errors.append(f"{master_dir.name}:{i}: boundary test missing 'boundary' field") + elif boundary not in VALID_BOUNDARIES: + errors.append( + f"{master_dir.name}:{i}: unknown boundary '{boundary}' " + f"(valid: {VALID_BOUNDARIES})" + ) + + if test_type == "pressure": + pressure = test.get("pressure") + if not pressure: + errors.append(f"{master_dir.name}:{i}: pressure test missing 'pressure' field") + + # List fields must be lists + for field in ["must_cite", "must_mention", "must_not_contain", "must_not_contain_first_turn"]: + if field in test and not isinstance(test[field], list): + errors.append(f"{master_dir.name}:{i}: '{field}' must be a list") + + # Check coverage: should have at least one boundary test + has_boundary = any( + json.loads(l).get("test_type") == "boundary" + for l in lines + if l.strip() + ) + if not has_boundary: + errors.append(f"{master_dir.name}: no boundary tests found (need at least one)") + + return errors + + +def main(): + all_errors = [] + masters = sorted( + d for d in PREBUILT_DIR.iterdir() + if d.is_dir() and (d / "tests" / "fidelity.jsonl").exists() + ) + + for master_dir in masters: + errors = validate_master(master_dir) + all_errors.extend(errors) + if not errors: + fidelity_path = master_dir / "tests" / "fidelity.jsonl" + count = len(fidelity_path.read_text().strip().splitlines()) if fidelity_path.exists() else 0 + print(f" {master_dir.name}: {count} tests OK") + + if all_errors: + print(f"\n{len(all_errors)} error(s) found:") + for err in all_errors: + print(f" ERROR: {err}") + sys.exit(1) + else: + print(f"\nAll {len(masters)} masters validated successfully.") + + +if __name__ == "__main__": + main()