# The AI Trust & Fairness Compendium validant.ai (Glinz & Company GmbH). Version 1.1 of 2026-11-06. Public. Creative Commons Attribution-NonCommercial 4.0 International: https://creativecommons.org/licenses/by-nc/4.0/. Canonical page: https://validant.ai/compendium Trust is assessed, not asserted. 12 parts + 115 chapters + 3 appendix volumes = one notation, one vocabulary. The compendium makes AI fairness checkable: twelve standalone parts with fixed chapter numbers and three appendix volumes, written from identified sources, reviewed and fact-checked, and read by part, by task or by question. ## The work in figures - Chapters: 115, fixed numbers across the whole work - Parts: 12 standalone documents - Pages: 2750 in 19 printed volumes - Defined terms: 933 (Appendix A, the controlled glossary) - Checked before publication: 2268 reference entries looked up, 2867 claims and 861 formulas rechecked ## Why this compendium AI systems already support decisions about who is hired, who receives credit and who gains access to services. The same technique faces different evidence, notice, review and audit obligations in credit, employment, insurance, healthcare, education and public services (Chapter 91). When such systems are not built and governed with care, they can carry existing social hierarchies forward: historical discrimination reaches contemporary AI through identifiable mechanisms such as biased targets and labels, proxies and feedback loops (Chapters 6 and 8). The difficulty is not a lack of fairness metrics. It is that fairness is a family of contested conceptions, that formal criteria encode different ones of them (Chapter 3), and that, when base rates differ between groups and prediction is imperfect, the main criteria cannot all hold at once (Chapter 19). Choosing among them is therefore unavoidable and normative (Chapter 22). A fairness metric computed without a stated question, a fixed analysis plan, validated measurements and quantified uncertainty is a number, not evidence (Chapter 36). And an assurance claim built on such evidence holds only for a stated scope, depth, criteria and date (Chapter 101). The compendium was written to close the gap between these insights and daily practice. It brings into one place, with one notation and one vocabulary, the knowledge an organization needs to examine an AI system for unfair treatment of people, to decide what to do about it, to document the decision and to show others that the claim it makes is supported. The chapters this text cites: Chapter 91: Sector-Specific Obligations; Chapter 6: Historical Patterns of Discrimination in Technology; Chapter 8: Modern Manifestations of Historical Biases; Chapter 3: Conceptions of Fairness and Justice; Chapter 19: Impossibility Theorems; Chapter 22: Contextual Selection of Fairness Criteria; Chapter 36: Fairness Assessment Methodology; Chapter 101: Limits of Certification and Assurance. ## Where to start for common tasks | Task | Start with | Continue with | | --- | --- | --- | | Start a fairness assessment of a system | Chapter 36 | Chapter 110, and its four methods | | Identify historical risks for a use case | Chapter 11 | Chapter 6, Chapter 7, Chapter 8 | | Choose fairness criteria | Chapter 22 | Chapters 14 to 20 | | Locate where bias enters | Chapter 35 | Chapter 25, Chapter 26, Chapter 27, Chapter 28, Chapter 29 | | Audit data before training | Chapter 37 and Chapter 38 | Chapter 24, Chapter 43 | | Choose and specify metrics | Chapter 44 | Chapter 39, Chapter 40, Chapter 41, Chapter 42, Chapter 43, and Appendix C | | Report a quantitative finding | Chapter 45 | Chapter 41, Chapter 42 | | Choose an intervention | Chapter 46 | Chapters 47 to 57, by intervention point | | Fix a model that cannot be retrained | Chapter 55 and Chapter 56 and Chapter 57 | Chapter 46 | | Automate fairness checks | Chapter 59 and Chapter 60 | Chapter 58 | | Monitor a deployed system | Chapter 61 and Chapter 62 | Chapter 63, Chapter 64 | | Assess a large language model or agent | Chapter 67 and Chapter 69 | Chapter 112, Chapter 113 | | Use explanations in an assessment | Chapter 76 and Chapter 77 | Chapter 71, Chapter 81 | | Meet a legal duty to explain | Chapter 79 | Chapter 86 | | Decide which obligations apply | Chapter 90 and Chapter 91 | Chapter 85, Chapter 86, Chapter 87, Chapter 88 | | Make a claim others can rely on | Chapter 93 | Chapter 94, Chapter 95, Chapter 96 | | Judge a certificate or audit report | Chapter 101 | Chapter 97, Chapter 98, Chapter 99, Chapter 100 | | Organize fairness work in a team | Chapter 102 | Chapter 103, Chapter 104, Chapter 105 | | Set up fairness governance | Chapter 106 | Chapter 107, Chapter 108, Chapter 109 | ## The questions the compendium answers - How do historical patterns of discrimination appear in technology, and which apply to a given system? Chapters 6, 7, 8, 11. - How are abstract ideas of fairness turned into precise definitions, and why do they conflict? Chapters 3, 12, 13, 14, 15, 16, 17, 18, 19, 20. - Where and how does bias enter the machine learning life cycle? Chapters 25, 26, 27, 28, 29, 35. - How is fairness quantified, validated and communicated? Chapters 39, 40, 41, 42, 43, 44. - How do fairness definitions translate into technical interventions? Chapters 30, 31, 32, 46. - Where in the pipeline should one intervene? Chapters 46. - Which technical approaches address which types of bias? Chapters 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57. - How are fairness improvements balanced against performance? Chapters 20, 52, 54. - How do existing fairness libraries differ, and how are they combined? Chapters 58. - Why are fairness metrics hard to compute correctly at scale? Chapters 40, 43. - Where in data pipelines should bias detection and mitigation occur? Chapters 38, 47, 48, 49, 60. - What threatens fairness after deployment, and how is it monitored? Chapters 34, 61, 62, 63, 64. - What do complex architectures such as language, ranking and vision models require? Chapters 65, 66, 67, 68, 69. ## Readers - Data scientists and ML engineers: Part II, Part III, Part IV, Part V, Part VI, Part VII, Part VIII; the front matter and Appendix 1 are written for every reader. - Auditors and assessors: Part I, Part II, Part III, Part IV, Part VII, Part VIII, Part X, Part XI; the front matter and Appendix 1 are written for every reader. - Risk, compliance and legal: Part III, Part VIII, Part IX, Part X, Part XI; the front matter and Appendix 1 are written for every reader. - Product owners and managers: Part I, Part II, Part IX, Part XI; the front matter and Appendix 1 are written for every reader. - Researchers: Part I, Part XII; the front matter and Appendix 1 are written for every reader. ## Contents ### Front matter: Front Matter and Master Table of Contents The way in: preface, executive overview of every part, scope and limitations, the editorial method and reading paths by task, with the full table of contents. PDF, 38 pages: https://validant.ai/compendium/pdf/Validant_Compendium_Front_Matter_and_Contents.pdf ### Part I: Foundations of Trust, Justice and Non-Discrimination Trust and appropriate reliance, conceptions of justice, equality and dignity, the history of discrimination in technology, and the Historical Context Assessment Method. Written for Auditors and assessors, Product owners and managers, Researchers. Version 1.0 of 2026-11-05, 181 pages. - Full document (PDF, 181 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_I_Foundations.pdf 1. The Nature of Trust 2. Trust in Artificial Intelligence 3. Conceptions of Fairness and Justice 4. Ethical Frameworks for Fairness Evaluation 5. Equality, Equity and Human Dignity 6. Historical Patterns of Discrimination in Technology 7. Data Representation and Social Context 8. Modern Manifestations of Historical Biases 9. Human Rights, Human Agency and Contestability 10. Fairness, Explainability and Accountability 11. The Historical Context Assessment Method ### Part II: Formal Foundations of Algorithmic Fairness The notation and the group, individual and causal fairness criteria, the impossibility theorems and trade-offs, and how to choose a criterion for a context. Written for Data scientists and ML engineers, Auditors and assessors, Product owners and managers. Version 1.4 of 2026-11-05, 185 pages. - Full document (PDF, 185 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_II_Formal_Foundations.pdf 12. Formalizing Algorithmic Decision-Making 13. Probability and Statistical Foundations 14. Independence: Demographic Parity and Related Criteria 15. Separation: Equal Opportunity and Equalized Odds 16. Sufficiency: Predictive Parity and Calibration 17. Individual Fairness 18. Causal and Counterfactual Fairness Criteria 19. Impossibility Theorems 20. Fairness Trade-offs 21. Fairness Beyond Binary Classification: Scores, Regression and Ranking 22. Contextual Selection of Fairness Criteria ### Part III: Bias, Discrimination and Causality Legal forms of discrimination, proxies, where bias enters the AI life cycle, intersectionality, causal models and counterfactual fairness, and distribution shift. Written for Data scientists and ML engineers, Auditors and assessors, Risk, compliance and legal. Version 1.2 of 2026-11-05, 258 pages in 2 volumes. - Volume 1: Chapters 23 to 29 (PDF, 139 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol1.pdf - Volume 2: Chapters 30 to 35 (PDF, 119 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol2.pdf 23. Direct, Indirect and Systemic Discrimination 24. Protected Characteristics, Sensitive Attributes and Proxies 25. Data Collection and Representation Biases 26. Algorithm Design and Implementation Biases 27. Feedback Loops and Amplification Effects 28. System Interaction and Deployment Biases 29. Intersectionality and Compound Discrimination 30. From Correlation to Causation 31. Building Causal Models for Bias Detection 32. The Counterfactual Fairness Framework 33. Causal Inference with Limited Data 34. Distribution Shift and Emerging Discrimination 35. The Bias Source Identification Method ### Part IV: Measuring Fairness: Metrics and Statistical Evidence Assessment methodology, data audits, group metrics and their implementation, confidence intervals and robustness, small subgroups, and how to select metrics. Written for Data scientists and ML engineers, Auditors and assessors. Version 1.1 of 2026-11-05, 200 pages. - Full document (PDF, 200 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_IV_Measuring_Fairness.pdf 36. Fairness Assessment Methodology 37. Comprehensive Data Auditing 38. Bias Detection in Raw Data 39. Group Fairness Metrics 40. Implementing Fairness Metrics 41. Statistical Validation and Confidence Intervals 42. Statistical Significance and Robustness 43. Small Samples and Sparse Subgroups 44. The Fairness Metrics Selection Method 45. Limits of Quantitative Fairness Assessment ### Part V: Fairness Interventions Choosing where to intervene, then pre-processing, in-processing and post-processing methods, each with its cost measured on the system at hand. Written for Data scientists and ML engineers. Version 1.1 of 2026-11-05, 251 pages in 2 volumes. - Volume 1: Chapters 46 to 52 (PDF, 141 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_V_Interventions_Vol1.pdf - Volume 2: Chapters 53 to 57 (PDF, 110 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_V_Interventions_Vol2.pdf 46. Choosing an Intervention Strategy 47. Reweighting and Resampling 48. Distribution Transformation 49. Feature Engineering and Transformation for Fairness 50. Fairness-Aware Data Generation 51. Fairness Objectives and Constraints 52. Fairness-Aware Loss Functions and Regularization 53. Adversarial Debiasing 54. Multi-Objective Optimization 55. Threshold Optimization 56. Calibration Across Groups 57. Prediction Transformation Methods ### Part VI: Fairness Engineering and Operations Fairness tools and libraries, checks in development and CI/CD, monitoring and drift alerts in production, dashboards, and A/B tests for fairness. Written for Data scientists and ML engineers. Version 1.1 of 2026-11-05, 135 pages. - Full document (PDF, 135 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_VI_Engineering_Operations.pdf 58. The Fairness Tooling Ecosystem 59. Fairness in the Development Workflow 60. Automated Bias Checks in CI/CD 61. Metric Tracking and Real-Time Bias Detection 62. Drift Detection and Alerting 63. Fairness Dashboards and Reporting 64. A/B Testing for Fairness ### Part VII: Fairness in Advanced AI Architectures Fairness in deep learning, recommendation and ranking, large language models, vision and multi-modal models, and generative and agentic systems. Written for Data scientists and ML engineers, Auditors and assessors. Version 1.0 of 2026-11-05, 107 pages. - Full document (PDF, 107 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_VII_Advanced_Architectures.pdf 65. Deep Learning Systems 66. Recommendation and Ranking Systems 67. Large Language Models 68. Multi-Modal and Vision Models 69. Generative and Agentic Systems ### Part VIII: Explainability, Interpretability and Non-Discrimination Explanation methods from attribution to counterfactuals and causal accounts, how to test their quality by group, legal rights to explanation, and their limits. Written for Data scientists and ML engineers, Auditors and assessors, Risk, compliance and legal. Version 1.0 of 2026-11-05, 246 pages in 2 volumes. - Volume 1: Chapters 70 to 76 (PDF, 140 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_VIII_Explainability_Vol1.pdf - Volume 2: Chapters 77 to 82 (PDF, 106 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_VIII_Explainability_Vol2.pdf 70. Foundations of AI Explainability 71. Explainability and Fairness 72. Taxonomy of Explainability Methods 73. Feature Attribution Methods 74. Counterfactual Explanations and Algorithmic Recourse 75. Causal Explainability 76. Group-Level and Intersectional Explanations 77. Explanation Quality, Fidelity and Robustness 78. Human-Centered Explainability 79. Explainability and Legal Rights 80. Explainability in Generative and Agentic AI 81. Explainability as Assurance Evidence 82. Limits, Risks and Fairness of Explanations ### Part IX: Legal, Ethical and Regulatory Foundations Non-discrimination law, human rights, the EU AI Act and GDPR Article 22, Swiss and US frameworks, standards, risk classification and sector duties. Written for Risk, compliance and legal, Product owners and managers. Version 1.0 of 2026-11-05, 195 pages. - Full document (PDF, 195 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_IX_Legal_Regulatory.pdf 83. Legal Standards for Algorithmic Fairness 84. International Human Rights and Non-Discrimination Law 85. The Global Regulatory Landscape 86. The EU AI Act and GDPR Article 22 87. Swiss Legal and Regulatory Framework 88. United States Anti-Discrimination Frameworks 89. International Standards and Governance Principles 90. Risk Classification and Assessment Systems 91. Sector-Specific Obligations 92. Ethics Beyond Compliance and Regulatory Change ### Part X: AI Assurance, Evidence and Digital Trust From measurement to assurance: arguments, evidence and audit trails, independent assessment, verifiable credentials, revocation, trust registries and the limits of certification. Written for Auditors and assessors, Risk, compliance and legal. Version 1.0 of 2026-11-05, 176 pages. - Full document (PDF, 176 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_X_Assurance_Digital_Trust.pdf 93. From Fairness Measurement to Assurance 94. Claims, Evidence and Assurance Arguments 95. Evidence Collection and Audit Trails 96. Independent Assessment and Reproducibility 97. Verifiable Credentials and Digital Attestations 98. Identity, Provenance and Evidence Integrity 99. Validity, Revocation and Continuous Assurance 100. Trust Registries and Distributed Trust Infrastructures 101. Limits of Certification and Assurance ### Part XI: Fairness in Organizational Practice Fairness in agile delivery, roles, documentation, governance, incident handling and redress, and the Fairness Audit Playbook with applied case studies. Written for Product owners and managers, Risk, compliance and legal, Auditors and assessors. Version 1.0 of 2026-11-05, 186 pages. - Full document (PDF, 186 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_XI_Organizational_Practice.pdf 102. Integrating Fairness into Agile Principles 103. Fairness User Stories and Acceptance Criteria 104. Sprint Planning and Execution for Fairness 105. Fairness Ceremonies and Checkpoints 106. Roles and Responsibilities for Fairness 107. Documentation and Communication Frameworks 108. Governance Mechanisms and Decision Processes 109. Incident Management, Remediation and Redress 110. The Fairness Audit Playbook: an Integrated Workflow 111. Applied Case Studies ### Part XII: Emerging Challenges and Research Frontiers Open frontiers: generative AI, autonomous agents and delegated decisions, adaptive systems and collective impacts, and an integrated research agenda. Written for Researchers. Version 1.0 of 2026-11-05, 77 pages. - Full document (PDF, 77 pages): https://validant.ai/compendium/pdf/Validant_Compendium_Part_XII_Research_Frontiers.pdf 112. Fairness in Large Language Models and Generative AI 113. Autonomous Agents and Delegated Decisions 114. Adaptive Systems and Collective Impacts 115. An Integrated Research Agenda ### Appendix Volume 1: Glossary, Notation and Index The controlled glossary (Appendix A), the mathematical notation (B) and the index of every definition with its chapters (L). PDF, 139 pages: https://validant.ai/compendium/pdf/Validant_Compendium_Appendices_Vol1_Glossary_Notation_Index.pdf ### Appendix Volume 2: Metrics, Methods and Worked Examples The catalogs of fairness metrics (C) and explainability methods (D), the statistical methods (E) and worked numerical examples to recompute (H). PDF, 158 pages: https://validant.ai/compendium/pdf/Validant_Compendium_Appendices_Vol2_Metrics_Methods_Examples.pdf ### Appendix Volume 3: Law, Standards, Templates, Evidence and Bibliography The legal and regulatory matrix (F), standards (G), method worksheets and templates (I), the evidence taxonomy (J) and the bibliography of every source (K). PDF, 218 pages: https://validant.ai/compendium/pdf/Validant_Compendium_Appendices_Vol3_Law_Standards_Templates_Bibliography.pdf ## For machines - Index for language models: https://validant.ai/compendium/llms.txt - Structure as data: https://validant.ai/compendium/index.json - Controlled glossary: https://validant.ai/compendium/glossary.json - Terms for crawlers (Really Simple Licensing): https://validant.ai/license.xml ## Licence The AI Trust & Fairness Compendium, validant.ai (Glinz & Company GmbH), is published under Creative Commons Attribution-NonCommercial 4.0 International (https://creativecommons.org/licenses/by-nc/4.0/). Read it, quote it, build on it for non-commercial work, and name the chapter, the part and the version you used. Commercial use, including training commercial models on the text, needs a word with hello@validant.ai. The compendium informs professional judgment; it is not legal advice for a specific case and does not certify any AI system. Law is stated as of November 2026. Downloads: 5 files per visitor are free (the front matter is not counted); an email address at https://validant.ai/compendium unlocks every file.