File contents IO_CONTRACT
input : 待证伪的技能假设 — 声明、适用场景、可检验条件
input : 真实任务输入(非 mock/synthetic)— 用于执行证伪测试
output : 证伪测试结果 — 每个测试含真实输入、期望输出、实际输出、通过/反证判定
output : 技能信任分级 — 0.9-1.0 继续使用 / 0.5-0.7 需重设计 / <0.5 替换技能
原则 (Principles)
以真为试 :证伪之试必用真实任务输入,mock 数据所成之验,不足以证其能。
留痕存证 :每试须存输入、过程日志、输出、度量于 test-results/,无迹之验不可复,不可复则不诚。
逆证为要 :但跑期望通过之试者,自欺也;反证之证与成功之证同须采集。
信由渐积 :信任分级依屡次证据而移,一单点之验不足以定高下,亦不足以成弃。
|
| 0.9-1.0 | High Trust | Continue using |
| 0.7-0.9 | Medium Trust | Increase monitoring |
| 0.5-0.7 | Low Trust | Redesign needed |
| < 0.5 | Unreliable | Replace skill |
Genes (策略基因)
紧凑策略表示。条件→策略。需要深度时参考完整文档。
[FALS-001] 执行证伪测试时 → 必须使用真实任务输入,严禁使用 mock 或合成数据
[FALS-002] 设计测试用例时 → 必须包含期望失败的反证测试,避免仅运行预期通过的自欺式验证
[FALS-003] 记录测试结果时 → 必须将输入、过程日志、输出及度量完整归档至 test-results/ 以确保可复现性
[FALS-004] 评估技能信任度时 → 必须基于屡次累积的证据动态调整分级,禁止依据单点数据定论
[FALS-005] 判定测试有效性时 → 必须采用量化指标对比期望与实际输出,拒绝主观的“看起来不错”评估
[FALS-006] 处理验证失败时 → 必须记录具体原因、上下文及修复建议,确保错误信息具备恢复指引
Evidence Requirements
Every test must produce:
Input Data : Real task input (not mock/synthetic)
Processing Log : Intermediate results from skill execution
Output Results : Final output from the skill
Comparison : Analysis of how output matches expectations
Metric Calculations : Quantitative quality scores
Evidence Storage : All artifacts saved to test-results/ directory
Anti-Patterns to Avoid
Testing with mock data instead of real data
Subjective "looks good" assessments without metrics
Only running tests that you expect to pass
Updating trust based on single data points
Ignoring failure evidence
Using inconsistent metric thresholds across tests
Not collecting evidence for tests that pass
Assuming skill works because it "seems right"
Related
Bayesian probability theory
Karl Popper's falsificationism
Scientific method and hypothesis testing
ML model evaluation and validation
A/B testing and experimental design
验证清单 · VERIFICATION
约束规则 · RULES
输入约束 : 参数类型、范围、格式必须校验
输出约束 : 返回值结构、编码、命名必须一致
异常约束 : 错误信息必须包含上下文和恢复建议
安全约束 : 不执行未验证的任意代码,不暴露内部状态
违反规则的操作视为不安全,必须拒绝或隔离。
每项验证必须可执行、可记录、可复现。验证失败时记录原因和修复。
Golden 集合 · GOLDEN SET
Golden Input : 待证伪技能声明 — 如 "citation-verification 技能在 pima-crispdm 数据集上可检出 ≥95% 虚构引用",含适用场景与可检验条件。
Golden Output : 每个测试产出完整证据链 — 真实输入(非 mock)、处理日志、最终输出、期望对比、量化指标,全部归档至 test-results/;技能信任分级更新(如 0.85 → High Trust,继续监控)。
Golden Error : 仅用 mock 数据执行测试,3/3 通过即判 "技能可靠" → 诊断:违反"以真为试"原则,反证路径缺失;修复:改用真实任务输入重跑,并补充期望失败的反证测试用例,重新采集证据。
1 --- 2 name: falsification-validation 3 description: Every test must produce: 4 license: MIT 5 --- 6 7 8 ## IO_CONTRACT 9 10 - **input**: 待证伪的技能假设 — 声明、适用场景、可检验条件 11 - **input**: 真实任务输入(非 mock/synthetic)— 用于执行证伪测试 12 - **output**: 证伪测试结果 — 每个测试含真实输入、期望输出、实际输出、通过/反证判定 13 - **output**: 技能信任分级 — 0.9-1.0 继续使用 / 0.5-0.7 需重设计 / <0.5 替换技能 14 15 ## 原则 (Principles) 16 17 - **以真为试**:证伪之试必用真实任务输入,mock 数据所成之验,不足以证其能。 18 - **留痕存证**:每试须存输入、过程日志、输出、度量于 `test-results/`,无迹之验不可复,不可复则不诚。 19 - **逆证为要**:但跑期望通过之试者,自欺也;反证之证与成功之证同须采集。 20 - **信由渐积**:信任分级依屡次证据而移,一单点之验不足以定高下,亦不足以成弃。 21 22 | 23 | 0.9-1.0 | High Trust | Continue using | 24 | 0.7-0.9 | Medium Trust | Increase monitoring | 25 | 0.5-0.7 | Low Trust | Redesign needed | 26 | < 0.5 | Unreliable | Replace skill | 27 28 29 ## Genes (策略基因) 30 31 > 紧凑策略表示。条件→策略。需要深度时参考完整文档。 32 33 - **[FALS-001]** 执行证伪测试时 → 必须使用真实任务输入,严禁使用 mock 或合成数据 34 - **[FALS-002]** 设计测试用例时 → 必须包含期望失败的反证测试,避免仅运行预期通过的自欺式验证 35 - **[FALS-003]** 记录测试结果时 → 必须将输入、过程日志、输出及度量完整归档至 `test-results/` 以确保可复现性 36 - **[FALS-004]** 评估技能信任度时 → 必须基于屡次累积的证据动态调整分级,禁止依据单点数据定论 37 - **[FALS-005]** 判定测试有效性时 → 必须采用量化指标对比期望与实际输出,拒绝主观的“看起来不错”评估 38 - **[FALS-006]** 处理验证失败时 → 必须记录具体原因、上下文及修复建议,确保错误信息具备恢复指引 39 40 ## Evidence Requirements 41 42 Every test must produce: 43 44 1. **Input Data**: Real task input (not mock/synthetic) 45 2. **Processing Log**: Intermediate results from skill execution 46 3. **Output Results**: Final output from the skill 47 4. **Comparison**: Analysis of how output matches expectations 48 5. **Metric Calculations**: Quantitative quality scores 49 6. **Evidence Storage**: All artifacts saved to `test-results/` directory 50 51 ## Anti-Patterns to Avoid 52 53 - Testing with mock data instead of real data 54 - Subjective "looks good" assessments without metrics 55 - Only running tests that you expect to pass 56 - Updating trust based on single data points 57 - Ignoring failure evidence 58 - Using inconsistent metric thresholds across tests 59 - Not collecting evidence for tests that pass 60 - Assuming skill works because it "seems right" 61 62 ## Related 63 64 - Bayesian probability theory 65 - Karl Popper's falsificationism 66 - Scientific method and hypothesis testing 67 - ML model evaluation and validation 68 - A/B testing and experimental design 69 70 ## 验证清单 · VERIFICATION 71 72 - [ ] 每个证伪测试均使用真实任务输入,无 mock/synthetic 数据(FALS-001、以真为试) 73 - [ ] 测试用例集包含期望失败的反证测试,未只跑预期通过的自欺式验证(FALS-002、逆证为要) 74 - [ ] 每项测试归档完整证据链至 `test-results/`:真实输入、过程日志、输出、期望对比、量化度量(FALS-003、Evidence Requirements 1-6) 75 - [ ] 测试判定采用量化指标对比期望与实际输出,无"看起来不错"式主观评估(FALS-005) 76 - [ ] 技能信任分级(0.9-1.0 / 0.7-0.9 / 0.5-0.7 / <0.5)基于屡次累积证据动态更新,未依据单点数据定论(FALS-004、信由渐积) 77 - [ ] 失败测试已记录具体原因、上下文与修复建议;通过测试同样留痕存证(FALS-006、Anti-Patterns: 不忽略失败证据、不为通过测试收集证据) 78 79 ## 约束规则 · RULES 80 81 1. **输入约束**: 参数类型、范围、格式必须校验 82 2. **输出约束**: 返回值结构、编码、命名必须一致 83 3. **异常约束**: 错误信息必须包含上下文和恢复建议 84 4. **安全约束**: 不执行未验证的任意代码,不暴露内部状态 85 86 > 违反规则的操作视为不安全,必须拒绝或隔离。 87 88 > 每项验证必须可执行、可记录、可复现。验证失败时记录原因和修复。 89 90 ## Golden 集合 · GOLDEN SET 91 92 - **Golden Input**: 待证伪技能声明 — 如 "citation-verification 技能在 pima-crispdm 数据集上可检出 ≥95% 虚构引用",含适用场景与可检验条件。 93 - **Golden Output**: 每个测试产出完整证据链 — 真实输入(非 mock)、处理日志、最终输出、期望对比、量化指标,全部归档至 `test-results/`;技能信任分级更新(如 0.85 → High Trust,继续监控)。 94 - **Golden Error**: 仅用 mock 数据执行测试,3/3 通过即判 "技能可靠" → 诊断:违反"以真为试"原则,反证路径缺失;修复:改用真实任务输入重跑,并补充期望失败的反证测试用例,重新采集证据。
yakeworld/Synthos/tree/main/skills/private/extended/research-tools/synthos-akne-bridge/quality/falsification-validation commit 09bcf84e20
Frequently asked questions How do I install the Falsification Validation skill? Run npx skillmds@latest add yakeworld/falsification-validation in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.
What does the Falsification Validation skill do? Every test must produce: It is listed under Research & Search on SkillMD.
Is Falsification Validation safe to use? This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.
Which AI agents work with Falsification Validation? This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.
Is Falsification Validation free to use? Yes. Installing skills from SkillMD is free. This skill is licensed under MIT.
Who published Falsification Validation? yakeworld (@yakeworld) published this skill. Their other Agent Skills are listed on their SkillMD profile.