{ "title": "The AI Trust & Fairness Compendium", "publisher": "validant.ai (Glinz & Company GmbH)", "url": "https://validant.ai/compendium", "version": "1.1", "date": "2026-11-06", "classification": "Public", "license": { "name": "Creative Commons Attribution-NonCommercial 4.0 International", "url": "https://creativecommons.org/licenses/by-nc/4.0/" }, "machineTerms": "https://validant.ai/license.xml", "llms": "https://validant.ai/compendium/llms.txt", "markdown": "https://validant.ai/compendium/index.md", "counts": { "parts": 12, "chapters": 115, "pages": 2750, "glossaryTerms": 933, "checked": { "references": 2268, "claims": 2867, "formulas": 861 } }, "readers": { "data-scientist": "Data scientists and ML engineers", "auditor": "Auditors and assessors", "compliance": "Risk, compliance and legal", "product": "Product owners and managers", "researcher": "Researchers" }, "parts": [ { "n": 1, "numeral": "I", "title": "Foundations of Trust, Justice and Non-Discrimination", "chapters": [ 1, 11 ], "document": { "id": "P01", "version": "1.0", "date": "2026-11-05", "pages": 181, "about": "Trust and appropriate reliance, conceptions of justice, equality and dignity, the history of discrimination in technology, and the Historical Context Assessment Method.", "keyMessage": "Fairness in AI is a normative and historical question before a technical one: trust must follow evidence of trustworthiness, the choice of fairness conception must be explicit, and documented past discrimination shows which risks to test for.", "audience": [ "auditor", "product", "researcher" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_I_Foundations.pdf", "pages": 181, "bytes": 4747013, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_I_Foundations.pdf" } ] } }, { "n": 2, "numeral": "II", "title": "Formal Foundations of Algorithmic Fairness", "chapters": [ 12, 22 ], "document": { "id": "P02", "version": "1.4", "date": "2026-11-05", "pages": 185, "about": "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.", "keyMessage": "Fairness criteria are precise statistical statements about predictions, outcomes and groups; the mathematics shows which of them can hold together, but choosing among them remains a normative decision.", "audience": [ "data-scientist", "auditor", "product" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_II_Formal_Foundations.pdf", "pages": 185, "bytes": 4719643, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_II_Formal_Foundations.pdf" } ] } }, { "n": 3, "numeral": "III", "title": "Bias, Discrimination and Causality", "chapters": [ 23, 35 ], "document": { "id": "P03", "version": "1.2", "date": "2026-11-05", "pages": 258, "about": "Legal forms of discrimination, proxies, where bias enters the AI life cycle, intersectionality, causal models and counterfactual fairness, and distribution shift.", "keyMessage": "Bias enters an AI system at identifiable points, from the data and the model to deployment and later shift; group metrics show that a disparity exists, while causal models explain why and where an intervention can remove it.", "audience": [ "data-scientist", "auditor", "compliance" ], "files": [ { "label": "Volume 1: Chapters 23 to 29", "file": "Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol1.pdf", "pages": 139, "bytes": 4119443, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol1.pdf" }, { "label": "Volume 2: Chapters 30 to 35", "file": "Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol2.pdf", "pages": 119, "bytes": 3743845, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol2.pdf" } ] } }, { "n": 4, "numeral": "IV", "title": "Measuring Fairness: Metrics and Statistical Evidence", "chapters": [ 36, 45 ], "document": { "id": "P04", "version": "1.1", "date": "2026-11-05", "pages": 200, "about": "Assessment methodology, data audits, group metrics and their implementation, confidence intervals and robustness, small subgroups, and how to select metrics.", "keyMessage": "A fairness metric becomes evidence only when it is computed correctly, reported with its denominators and uncertainty, tested for robustness across every group including small ones, and scoped to what quantitative measurement can establish.", "audience": [ "data-scientist", "auditor" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_IV_Measuring_Fairness.pdf", "pages": 200, "bytes": 4914874, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_IV_Measuring_Fairness.pdf" } ] } }, { "n": 5, "numeral": "V", "title": "Fairness Interventions", "chapters": [ 46, 57 ], "document": { "id": "P05", "version": "1.1", "date": "2026-11-05", "pages": 251, "about": "Choosing where to intervene, then pre-processing, in-processing and post-processing methods, each with its cost measured on the system at hand.", "keyMessage": "Fairness interventions act on the training data, on training itself or on a model's outputs; the right point follows from where a disparity enters, and the price of each method must be measured on the specific system.", "audience": [ "data-scientist" ], "files": [ { "label": "Volume 1: Chapters 46 to 52", "file": "Validant_Compendium_Part_V_Interventions_Vol1.pdf", "pages": 141, "bytes": 4120019, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_V_Interventions_Vol1.pdf" }, { "label": "Volume 2: Chapters 53 to 57", "file": "Validant_Compendium_Part_V_Interventions_Vol2.pdf", "pages": 110, "bytes": 3514968, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_V_Interventions_Vol2.pdf" } ] } }, { "n": 6, "numeral": "VI", "title": "Fairness Engineering and Operations", "chapters": [ 58, 64 ], "document": { "id": "P06", "version": "1.1", "date": "2026-11-05", "pages": 135, "about": "Fairness tools and libraries, checks in development and CI/CD, monitoring and drift alerts in production, dashboards, and A/B tests for fairness.", "keyMessage": "Fairness becomes dependable only when it is executable: measured with well-chosen tools, tested automatically before release, monitored per group in production, reported to the people who decide, and changed only after controlled experiments.", "audience": [ "data-scientist" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_VI_Engineering_Operations.pdf", "pages": 135, "bytes": 4209553, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VI_Engineering_Operations.pdf" } ] } }, { "n": 7, "numeral": "VII", "title": "Fairness in Advanced AI Architectures", "chapters": [ 65, 69 ], "document": { "id": "P07", "version": "1.0", "date": "2026-11-05", "pages": 107, "about": "Fairness in deep learning, recommendation and ranking, large language models, vision and multi-modal models, and generative and agentic systems.", "keyMessage": "In advanced architectures unfairness lives in learned representations, exposure, generated content, fused modalities and multi-step actions, so fairness must be assessed inside the system, over time and across the ecosystem, not only on final decisions.", "audience": [ "data-scientist", "auditor" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_VII_Advanced_Architectures.pdf", "pages": 107, "bytes": 3017087, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VII_Advanced_Architectures.pdf" } ] } }, { "n": 8, "numeral": "VIII", "title": "Explainability, Interpretability and Non-Discrimination", "chapters": [ 70, 82 ], "document": { "id": "P08", "version": "1.0", "date": "2026-11-05", "pages": 246, "about": "Explanation methods from attribution to counterfactuals and causal accounts, how to test their quality by group, legal rights to explanation, and their limits.", "keyMessage": "Explanations can help detect, contest and remedy unfairness, but they are evidence only when faithful to the model, robust, tested by group and understood by their audience; they show how a model works, not that it is fair.", "audience": [ "data-scientist", "auditor", "compliance" ], "files": [ { "label": "Volume 1: Chapters 70 to 76", "file": "Validant_Compendium_Part_VIII_Explainability_Vol1.pdf", "pages": 140, "bytes": 4173631, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VIII_Explainability_Vol1.pdf" }, { "label": "Volume 2: Chapters 77 to 82", "file": "Validant_Compendium_Part_VIII_Explainability_Vol2.pdf", "pages": 106, "bytes": 3296893, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VIII_Explainability_Vol2.pdf" } ] } }, { "n": 9, "numeral": "IX", "title": "Legal, Ethical and Regulatory Foundations", "chapters": [ 83, 92 ], "document": { "id": "P09", "version": "1.0", "date": "2026-11-05", "pages": 195, "about": "Non-discrimination law, human rights, the EU AI Act and GDPR Article 22, Swiss and US frameworks, standards, risk classification and sector duties.", "keyMessage": "Law sets a mandatory but moving floor for AI fairness: it names protected grounds, defines direct and indirect discrimination and imposes risk-based and sectoral duties, yet no metric proves compliance and compliance alone does not establish fairness.", "audience": [ "compliance", "product" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_IX_Legal_Regulatory.pdf", "pages": 195, "bytes": 4784797, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_IX_Legal_Regulatory.pdf" } ] } }, { "n": 10, "numeral": "X", "title": "AI Assurance, Evidence and Digital Trust", "chapters": [ 93, 101 ], "document": { "id": "P10", "version": "1.0", "date": "2026-11-05", "pages": 176, "about": "From measurement to assurance: arguments, evidence and audit trails, independent assessment, verifiable credentials, revocation, trust registries and the limits of certification.", "keyMessage": "A fairness metric becomes assurance only inside a scoped, argued claim backed by preserved evidence and independent checking; credentials, registries and status checks make such claims verifiable over time, but no certificate shows that a system is fair in general.", "audience": [ "auditor", "compliance" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_X_Assurance_Digital_Trust.pdf", "pages": 176, "bytes": 4635649, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_X_Assurance_Digital_Trust.pdf" } ] } }, { "n": 11, "numeral": "XI", "title": "Fairness in Organizational Practice", "chapters": [ 102, 111 ], "document": { "id": "P11", "version": "1.0", "date": "2026-11-05", "pages": 186, "about": "Fairness in agile delivery, roles, documentation, governance, incident handling and redress, and the Fairness Audit Playbook with applied case studies.", "keyMessage": "Fairness becomes organizational practice only when delivery, roles, documentation, governance and incident handling carry explicit fairness content, so that bias work is planned, owned, recorded, decided and repaired rather than left to goodwill.", "audience": [ "product", "compliance", "auditor" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_XI_Organizational_Practice.pdf", "pages": 186, "bytes": 4889037, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_XI_Organizational_Practice.pdf" } ] } }, { "n": 12, "numeral": "XII", "title": "Emerging Challenges and Research Frontiers", "chapters": [ 112, 115 ], "document": { "id": "P12", "version": "1.0", "date": "2026-11-05", "pages": 77, "about": "Open frontiers: generative AI, autonomous agents and delegated decisions, adaptive systems and collective impacts, and an integrated research agenda.", "keyMessage": "At the frontier, fairness is no longer a property of one model's decisions but of evidence about open-ended outputs, delegated action and systems acting over time and across ecosystems; valid methods for producing that evidence are still being built.", "audience": [ "researcher" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_XII_Research_Frontiers.pdf", "pages": 77, "bytes": 2664893, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_XII_Research_Frontiers.pdf" } ] } } ], "chapters": [ { "n": 1, "title": "The Nature of Trust", "part": "I" }, { "n": 2, "title": "Trust in Artificial Intelligence", "part": "I" }, { "n": 3, "title": "Conceptions of Fairness and Justice", "part": "I" }, { "n": 4, "title": "Ethical Frameworks for Fairness Evaluation", "part": "I" }, { "n": 5, "title": "Equality, Equity and Human Dignity", "part": "I" }, { "n": 6, "title": "Historical Patterns of Discrimination in Technology", "part": "I" }, { "n": 7, "title": "Data Representation and Social Context", "part": "I" }, { "n": 8, "title": "Modern Manifestations of Historical Biases", "part": "I" }, { "n": 9, "title": "Human Rights, Human Agency and Contestability", "part": "I" }, { "n": 10, "title": "Fairness, Explainability and Accountability", "part": "I" }, { "n": 11, "title": "The Historical Context Assessment Method", "part": "I" }, { "n": 12, "title": "Formalizing Algorithmic Decision-Making", "part": "II" }, { "n": 13, "title": "Probability and Statistical Foundations", "part": "II" }, { "n": 14, "title": "Independence: Demographic Parity and Related Criteria", "part": "II" }, { "n": 15, "title": "Separation: Equal Opportunity and Equalized Odds", "part": "II" }, { "n": 16, "title": "Sufficiency: Predictive Parity and Calibration", "part": "II" }, { "n": 17, "title": "Individual Fairness", "part": "II" }, { "n": 18, "title": "Causal and Counterfactual Fairness Criteria", "part": "II" }, { "n": 19, "title": "Impossibility Theorems", "part": "II" }, { "n": 20, "title": "Fairness Trade-offs", "part": "II" }, { "n": 21, "title": "Fairness Beyond Binary Classification: Scores, Regression and Ranking", "part": "II" }, { "n": 22, "title": "Contextual Selection of Fairness Criteria", "part": "II" }, { "n": 23, "title": "Direct, Indirect and Systemic Discrimination", "part": "III" }, { "n": 24, "title": "Protected Characteristics, Sensitive Attributes and Proxies", "part": "III" }, { "n": 25, "title": "Data Collection and Representation Biases", "part": "III" }, { "n": 26, "title": "Algorithm Design and Implementation Biases", "part": "III" }, { "n": 27, "title": "Feedback Loops and Amplification Effects", "part": "III" }, { "n": 28, "title": "System Interaction and Deployment Biases", "part": "III" }, { "n": 29, "title": "Intersectionality and Compound Discrimination", "part": "III" }, { "n": 30, "title": "From Correlation to Causation", "part": "III" }, { "n": 31, "title": "Building Causal Models for Bias Detection", "part": "III" }, { "n": 32, "title": "The Counterfactual Fairness Framework", "part": "III" }, { "n": 33, "title": "Causal Inference with Limited Data", "part": "III" }, { "n": 34, "title": "Distribution Shift and Emerging Discrimination", "part": "III" }, { "n": 35, "title": "The Bias Source Identification Method", "part": "III" }, { "n": 36, "title": "Fairness Assessment Methodology", "part": "IV" }, { "n": 37, "title": "Comprehensive Data Auditing", "part": "IV" }, { "n": 38, "title": "Bias Detection in Raw Data", "part": "IV" }, { "n": 39, "title": "Group Fairness Metrics", "part": "IV" }, { "n": 40, "title": "Implementing Fairness Metrics", "part": "IV" }, { "n": 41, "title": "Statistical Validation and Confidence Intervals", "part": "IV" }, { "n": 42, "title": "Statistical Significance and Robustness", "part": "IV" }, { "n": 43, "title": "Small Samples and Sparse Subgroups", "part": "IV" }, { "n": 44, "title": "The Fairness Metrics Selection Method", "part": "IV" }, { "n": 45, "title": "Limits of Quantitative Fairness Assessment", "part": "IV" }, { "n": 46, "title": "Choosing an Intervention Strategy", "part": "V" }, { "n": 47, "title": "Reweighting and Resampling", "part": "V" }, { "n": 48, "title": "Distribution Transformation", "part": "V" }, { "n": 49, "title": "Feature Engineering and Transformation for Fairness", "part": "V" }, { "n": 50, "title": "Fairness-Aware Data Generation", "part": "V" }, { "n": 51, "title": "Fairness Objectives and Constraints", "part": "V" }, { "n": 52, "title": "Fairness-Aware Loss Functions and Regularization", "part": "V" }, { "n": 53, "title": "Adversarial Debiasing", "part": "V" }, { "n": 54, "title": "Multi-Objective Optimization", "part": "V" }, { "n": 55, "title": "Threshold Optimization", "part": "V" }, { "n": 56, "title": "Calibration Across Groups", "part": "V" }, { "n": 57, "title": "Prediction Transformation Methods", "part": "V" }, { "n": 58, "title": "The Fairness Tooling Ecosystem", "part": "VI" }, { "n": 59, "title": "Fairness in the Development Workflow", "part": "VI" }, { "n": 60, "title": "Automated Bias Checks in CI/CD", "part": "VI" }, { "n": 61, "title": "Metric Tracking and Real-Time Bias Detection", "part": "VI" }, { "n": 62, "title": "Drift Detection and Alerting", "part": "VI" }, { "n": 63, "title": "Fairness Dashboards and Reporting", "part": "VI" }, { "n": 64, "title": "A/B Testing for Fairness", "part": "VI" }, { "n": 65, "title": "Deep Learning Systems", "part": "VII" }, { "n": 66, "title": "Recommendation and Ranking Systems", "part": "VII" }, { "n": 67, "title": "Large Language Models", "part": "VII" }, { "n": 68, "title": "Multi-Modal and Vision Models", "part": "VII" }, { "n": 69, "title": "Generative and Agentic Systems", "part": "VII" }, { "n": 70, "title": "Foundations of AI Explainability", "part": "VIII" }, { "n": 71, "title": "Explainability and Fairness", "part": "VIII" }, { "n": 72, "title": "Taxonomy of Explainability Methods", "part": "VIII" }, { "n": 73, "title": "Feature Attribution Methods", "part": "VIII" }, { "n": 74, "title": "Counterfactual Explanations and Algorithmic Recourse", "part": "VIII" }, { "n": 75, "title": "Causal Explainability", "part": "VIII" }, { "n": 76, "title": "Group-Level and Intersectional Explanations", "part": "VIII" }, { "n": 77, "title": "Explanation Quality, Fidelity and Robustness", "part": "VIII" }, { "n": 78, "title": "Human-Centered Explainability", "part": "VIII" }, { "n": 79, "title": "Explainability and Legal Rights", "part": "VIII" }, { "n": 80, "title": "Explainability in Generative and Agentic AI", "part": "VIII" }, { "n": 81, "title": "Explainability as Assurance Evidence", "part": "VIII" }, { "n": 82, "title": "Limits, Risks and Fairness of Explanations", "part": "VIII" }, { "n": 83, "title": "Legal Standards for Algorithmic Fairness", "part": "IX" }, { "n": 84, "title": "International Human Rights and Non-Discrimination Law", "part": "IX" }, { "n": 85, "title": "The Global Regulatory Landscape", "part": "IX" }, { "n": 86, "title": "The EU AI Act and GDPR Article 22", "part": "IX" }, { "n": 87, "title": "Swiss Legal and Regulatory Framework", "part": "IX" }, { "n": 88, "title": "United States Anti-Discrimination Frameworks", "part": "IX" }, { "n": 89, "title": "International Standards and Governance Principles", "part": "IX" }, { "n": 90, "title": "Risk Classification and Assessment Systems", "part": "IX" }, { "n": 91, "title": "Sector-Specific Obligations", "part": "IX" }, { "n": 92, "title": "Ethics Beyond Compliance and Regulatory Change", "part": "IX" }, { "n": 93, "title": "From Fairness Measurement to Assurance", "part": "X" }, { "n": 94, "title": "Claims, Evidence and Assurance Arguments", "part": "X" }, { "n": 95, "title": "Evidence Collection and Audit Trails", "part": "X" }, { "n": 96, "title": "Independent Assessment and Reproducibility", "part": "X" }, { "n": 97, "title": "Verifiable Credentials and Digital Attestations", "part": "X" }, { "n": 98, "title": "Identity, Provenance and Evidence Integrity", "part": "X" }, { "n": 99, "title": "Validity, Revocation and Continuous Assurance", "part": "X" }, { "n": 100, "title": "Trust Registries and Distributed Trust Infrastructures", "part": "X" }, { "n": 101, "title": "Limits of Certification and Assurance", "part": "X" }, { "n": 102, "title": "Integrating Fairness into Agile Principles", "part": "XI" }, { "n": 103, "title": "Fairness User Stories and Acceptance Criteria", "part": "XI" }, { "n": 104, "title": "Sprint Planning and Execution for Fairness", "part": "XI" }, { "n": 105, "title": "Fairness Ceremonies and Checkpoints", "part": "XI" }, { "n": 106, "title": "Roles and Responsibilities for Fairness", "part": "XI" }, { "n": 107, "title": "Documentation and Communication Frameworks", "part": "XI" }, { "n": 108, "title": "Governance Mechanisms and Decision Processes", "part": "XI" }, { "n": 109, "title": "Incident Management, Remediation and Redress", "part": "XI" }, { "n": 110, "title": "The Fairness Audit Playbook: an Integrated Workflow", "part": "XI" }, { "n": 111, "title": "Applied Case Studies", "part": "XI" }, { "n": 112, "title": "Fairness in Large Language Models and Generative AI", "part": "XII" }, { "n": 113, "title": "Autonomous Agents and Delegated Decisions", "part": "XII" }, { "n": 114, "title": "Adaptive Systems and Collective Impacts", "part": "XII" }, { "n": 115, "title": "An Integrated Research Agenda", "part": "XII" } ], "readingPaths": [ { "task": "Start a fairness assessment of a system", "start": [ 36 ], "next": [ 110 ], "note": "and its four methods" }, { "task": "Identify historical risks for a use case", "start": [ 11 ], "next": [ 6, 7, 8 ] }, { "task": "Choose fairness criteria", "start": [ 22 ], "next": [ 14, 15, 16, 17, 18, 19, 20 ] }, { "task": "Locate where bias enters", "start": [ 35 ], "next": [ 25, 26, 27, 28, 29 ] }, { "task": "Audit data before training", "start": [ 37, 38 ], "next": [ 24, 43 ] }, { "task": "Choose and specify metrics", "start": [ 44 ], "next": [ 39, 40, 41, 42, 43 ], "note": "and Appendix C" }, { "task": "Report a quantitative finding", "start": [ 45 ], "next": [ 41, 42 ] }, { "task": "Choose an intervention", "start": [ 46 ], "next": [ 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57 ], "note": "by intervention point" }, { "task": "Fix a model that cannot be retrained", "start": [ 55, 56, 57 ], "next": [ 46 ] }, { "task": "Automate fairness checks", "start": [ 59, 60 ], "next": [ 58 ] }, { "task": "Monitor a deployed system", "start": [ 61, 62 ], "next": [ 63, 64 ] }, { "task": "Assess a large language model or agent", "start": [ 67, 69 ], "next": [ 112, 113 ] }, { "task": "Use explanations in an assessment", "start": [ 76, 77 ], "next": [ 71, 81 ] }, { "task": "Meet a legal duty to explain", "start": [ 79 ], "next": [ 86 ] }, { "task": "Decide which obligations apply", "start": [ 90, 91 ], "next": [ 85, 86, 87, 88 ] }, { "task": "Make a claim others can rely on", "start": [ 93 ], "next": [ 94, 95, 96 ] }, { "task": "Judge a certificate or audit report", "start": [ 101 ], "next": [ 97, 98, 99, 100 ] }, { "task": "Organize fairness work in a team", "start": [ 102 ], "next": [ 103, 104, 105 ] }, { "task": "Set up fairness governance", "start": [ 106 ], "next": [ 107, 108, 109 ] } ], "questions": [ { "q": "How do historical patterns of discrimination appear in technology, and which apply to a given system?", "ch": [ 6, 7, 8, 11 ] }, { "q": "How are abstract ideas of fairness turned into precise definitions, and why do they conflict?", "ch": [ 3, 12, 13, 14, 15, 16, 17, 18, 19, 20 ] }, { "q": "Where and how does bias enter the machine learning life cycle?", "ch": [ 25, 26, 27, 28, 29, 35 ] }, { "q": "How is fairness quantified, validated and communicated?", "ch": [ 39, 40, 41, 42, 43, 44 ] }, { "q": "How do fairness definitions translate into technical interventions?", "ch": [ 30, 31, 32, 46 ] }, { "q": "Where in the pipeline should one intervene?", "ch": [ 46 ] }, { "q": "Which technical approaches address which types of bias?", "ch": [ 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57 ] }, { "q": "How are fairness improvements balanced against performance?", "ch": [ 20, 52, 54 ] }, { "q": "How do existing fairness libraries differ, and how are they combined?", "ch": [ 58 ] }, { "q": "Why are fairness metrics hard to compute correctly at scale?", "ch": [ 40, 43 ] }, { "q": "Where in data pipelines should bias detection and mitigation occur?", "ch": [ 38, 47, 48, 49, 60 ] }, { "q": "What threatens fairness after deployment, and how is it monitored?", "ch": [ 34, 61, 62, 63, 64 ] }, { "q": "What do complex architectures such as language, ranking and vision models require?", "ch": [ 65, 66, 67, 68, 69 ] } ], "whyThisCompendium": { "chapters": [ 91, 6, 8, 3, 19, 22, 36, 101 ] }, "documents": [ { "id": "P00", "kind": "front", "numeral": "F", "label": "Front matter", "title": "Front Matter and Master Table of Contents", "docId": "VAI-COMP-P00", "version": "1.1", "date": "2026-11-06", "chapters": null, "chapterCount": 0, "about": "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.", "keyMessage": "The compendium makes AI fairness checkable: twelve standalone parts with 115 fixed chapter numbers and three appendix volumes, written from identified sources, reviewed and fact-checked, and read by part, by task or by question.", "audience": [ "data-scientist", "auditor", "compliance", "product", "researcher" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Front_Matter_and_Contents.pdf", "pages": 38, "bytes": 1283604, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Front_Matter_and_Contents.pdf" } ] }, { "id": "P01", "kind": "part", "numeral": "I", "label": "Part I", "title": "Foundations of Trust, Justice and Non-Discrimination", "docId": "VAI-COMP-P01", "version": "1.0", "date": "2026-11-05", "chapters": [ 1, 11 ], "chapterCount": 11, "about": "Trust and appropriate reliance, conceptions of justice, equality and dignity, the history of discrimination in technology, and the Historical Context Assessment Method.", "keyMessage": "Fairness in AI is a normative and historical question before a technical one: trust must follow evidence of trustworthiness, the choice of fairness conception must be explicit, and documented past discrimination shows which risks to test for.", "audience": [ "auditor", "product", "researcher" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_I_Foundations.pdf", "pages": 181, "bytes": 4747013, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_I_Foundations.pdf" } ] }, { "id": "P02", "kind": "part", "numeral": "II", "label": "Part II", "title": "Formal Foundations of Algorithmic Fairness", "docId": "VAI-COMP-P02", "version": "1.4", "date": "2026-11-05", "chapters": [ 12, 22 ], "chapterCount": 11, "about": "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.", "keyMessage": "Fairness criteria are precise statistical statements about predictions, outcomes and groups; the mathematics shows which of them can hold together, but choosing among them remains a normative decision.", "audience": [ "data-scientist", "auditor", "product" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_II_Formal_Foundations.pdf", "pages": 185, "bytes": 4719643, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_II_Formal_Foundations.pdf" } ] }, { "id": "P03", "kind": "part", "numeral": "III", "label": "Part III", "title": "Bias, Discrimination and Causality", "docId": "VAI-COMP-P03", "version": "1.2", "date": "2026-11-05", "chapters": [ 23, 35 ], "chapterCount": 13, "about": "Legal forms of discrimination, proxies, where bias enters the AI life cycle, intersectionality, causal models and counterfactual fairness, and distribution shift.", "keyMessage": "Bias enters an AI system at identifiable points, from the data and the model to deployment and later shift; group metrics show that a disparity exists, while causal models explain why and where an intervention can remove it.", "audience": [ "data-scientist", "auditor", "compliance" ], "files": [ { "label": "Volume 1: Chapters 23 to 29", "file": "Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol1.pdf", "pages": 139, "bytes": 4119443, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol1.pdf" }, { "label": "Volume 2: Chapters 30 to 35", "file": "Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol2.pdf", "pages": 119, "bytes": 3743845, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_III_Bias_Discrimination_Causality_Vol2.pdf" } ] }, { "id": "P04", "kind": "part", "numeral": "IV", "label": "Part IV", "title": "Measuring Fairness: Metrics and Statistical Evidence", "docId": "VAI-COMP-P04", "version": "1.1", "date": "2026-11-05", "chapters": [ 36, 45 ], "chapterCount": 10, "about": "Assessment methodology, data audits, group metrics and their implementation, confidence intervals and robustness, small subgroups, and how to select metrics.", "keyMessage": "A fairness metric becomes evidence only when it is computed correctly, reported with its denominators and uncertainty, tested for robustness across every group including small ones, and scoped to what quantitative measurement can establish.", "audience": [ "data-scientist", "auditor" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_IV_Measuring_Fairness.pdf", "pages": 200, "bytes": 4914874, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_IV_Measuring_Fairness.pdf" } ] }, { "id": "P05", "kind": "part", "numeral": "V", "label": "Part V", "title": "Fairness Interventions", "docId": "VAI-COMP-P05", "version": "1.1", "date": "2026-11-05", "chapters": [ 46, 57 ], "chapterCount": 12, "about": "Choosing where to intervene, then pre-processing, in-processing and post-processing methods, each with its cost measured on the system at hand.", "keyMessage": "Fairness interventions act on the training data, on training itself or on a model's outputs; the right point follows from where a disparity enters, and the price of each method must be measured on the specific system.", "audience": [ "data-scientist" ], "files": [ { "label": "Volume 1: Chapters 46 to 52", "file": "Validant_Compendium_Part_V_Interventions_Vol1.pdf", "pages": 141, "bytes": 4120019, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_V_Interventions_Vol1.pdf" }, { "label": "Volume 2: Chapters 53 to 57", "file": "Validant_Compendium_Part_V_Interventions_Vol2.pdf", "pages": 110, "bytes": 3514968, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_V_Interventions_Vol2.pdf" } ] }, { "id": "P06", "kind": "part", "numeral": "VI", "label": "Part VI", "title": "Fairness Engineering and Operations", "docId": "VAI-COMP-P06", "version": "1.1", "date": "2026-11-05", "chapters": [ 58, 64 ], "chapterCount": 7, "about": "Fairness tools and libraries, checks in development and CI/CD, monitoring and drift alerts in production, dashboards, and A/B tests for fairness.", "keyMessage": "Fairness becomes dependable only when it is executable: measured with well-chosen tools, tested automatically before release, monitored per group in production, reported to the people who decide, and changed only after controlled experiments.", "audience": [ "data-scientist" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_VI_Engineering_Operations.pdf", "pages": 135, "bytes": 4209553, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VI_Engineering_Operations.pdf" } ] }, { "id": "P07", "kind": "part", "numeral": "VII", "label": "Part VII", "title": "Fairness in Advanced AI Architectures", "docId": "VAI-COMP-P07", "version": "1.0", "date": "2026-11-05", "chapters": [ 65, 69 ], "chapterCount": 5, "about": "Fairness in deep learning, recommendation and ranking, large language models, vision and multi-modal models, and generative and agentic systems.", "keyMessage": "In advanced architectures unfairness lives in learned representations, exposure, generated content, fused modalities and multi-step actions, so fairness must be assessed inside the system, over time and across the ecosystem, not only on final decisions.", "audience": [ "data-scientist", "auditor" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_VII_Advanced_Architectures.pdf", "pages": 107, "bytes": 3017087, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VII_Advanced_Architectures.pdf" } ] }, { "id": "P08", "kind": "part", "numeral": "VIII", "label": "Part VIII", "title": "Explainability, Interpretability and Non-Discrimination", "docId": "VAI-COMP-P08", "version": "1.0", "date": "2026-11-05", "chapters": [ 70, 82 ], "chapterCount": 13, "about": "Explanation methods from attribution to counterfactuals and causal accounts, how to test their quality by group, legal rights to explanation, and their limits.", "keyMessage": "Explanations can help detect, contest and remedy unfairness, but they are evidence only when faithful to the model, robust, tested by group and understood by their audience; they show how a model works, not that it is fair.", "audience": [ "data-scientist", "auditor", "compliance" ], "files": [ { "label": "Volume 1: Chapters 70 to 76", "file": "Validant_Compendium_Part_VIII_Explainability_Vol1.pdf", "pages": 140, "bytes": 4173631, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VIII_Explainability_Vol1.pdf" }, { "label": "Volume 2: Chapters 77 to 82", "file": "Validant_Compendium_Part_VIII_Explainability_Vol2.pdf", "pages": 106, "bytes": 3296893, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_VIII_Explainability_Vol2.pdf" } ] }, { "id": "P09", "kind": "part", "numeral": "IX", "label": "Part IX", "title": "Legal, Ethical and Regulatory Foundations", "docId": "VAI-COMP-P09", "version": "1.0", "date": "2026-11-05", "chapters": [ 83, 92 ], "chapterCount": 10, "about": "Non-discrimination law, human rights, the EU AI Act and GDPR Article 22, Swiss and US frameworks, standards, risk classification and sector duties.", "keyMessage": "Law sets a mandatory but moving floor for AI fairness: it names protected grounds, defines direct and indirect discrimination and imposes risk-based and sectoral duties, yet no metric proves compliance and compliance alone does not establish fairness.", "audience": [ "compliance", "product" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_IX_Legal_Regulatory.pdf", "pages": 195, "bytes": 4784797, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_IX_Legal_Regulatory.pdf" } ] }, { "id": "P10", "kind": "part", "numeral": "X", "label": "Part X", "title": "AI Assurance, Evidence and Digital Trust", "docId": "VAI-COMP-P10", "version": "1.0", "date": "2026-11-05", "chapters": [ 93, 101 ], "chapterCount": 9, "about": "From measurement to assurance: arguments, evidence and audit trails, independent assessment, verifiable credentials, revocation, trust registries and the limits of certification.", "keyMessage": "A fairness metric becomes assurance only inside a scoped, argued claim backed by preserved evidence and independent checking; credentials, registries and status checks make such claims verifiable over time, but no certificate shows that a system is fair in general.", "audience": [ "auditor", "compliance" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_X_Assurance_Digital_Trust.pdf", "pages": 176, "bytes": 4635649, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_X_Assurance_Digital_Trust.pdf" } ] }, { "id": "P11", "kind": "part", "numeral": "XI", "label": "Part XI", "title": "Fairness in Organizational Practice", "docId": "VAI-COMP-P11", "version": "1.0", "date": "2026-11-05", "chapters": [ 102, 111 ], "chapterCount": 10, "about": "Fairness in agile delivery, roles, documentation, governance, incident handling and redress, and the Fairness Audit Playbook with applied case studies.", "keyMessage": "Fairness becomes organizational practice only when delivery, roles, documentation, governance and incident handling carry explicit fairness content, so that bias work is planned, owned, recorded, decided and repaired rather than left to goodwill.", "audience": [ "product", "compliance", "auditor" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_XI_Organizational_Practice.pdf", "pages": 186, "bytes": 4889037, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_XI_Organizational_Practice.pdf" } ] }, { "id": "P12", "kind": "part", "numeral": "XII", "label": "Part XII", "title": "Emerging Challenges and Research Frontiers", "docId": "VAI-COMP-P12", "version": "1.0", "date": "2026-11-05", "chapters": [ 112, 115 ], "chapterCount": 4, "about": "Open frontiers: generative AI, autonomous agents and delegated decisions, adaptive systems and collective impacts, and an integrated research agenda.", "keyMessage": "At the frontier, fairness is no longer a property of one model's decisions but of evidence about open-ended outputs, delegated action and systems acting over time and across ecosystems; valid methods for producing that evidence are still being built.", "audience": [ "researcher" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Part_XII_Research_Frontiers.pdf", "pages": 77, "bytes": 2664893, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Part_XII_Research_Frontiers.pdf" } ] }, { "id": "A1", "kind": "appendix", "numeral": "A1", "label": "Appendix Volume 1", "title": "Glossary, Notation and Index", "docId": "VAI-COMP-APP1", "version": "1.1", "date": "2026-11-05", "chapters": null, "chapterCount": 0, "about": "The controlled glossary (Appendix A), the mathematical notation (B) and the index of every definition with its chapters (L).", "keyMessage": "One vocabulary and one notation for the whole compendium: each defined term and each symbol is listed with the chapter that defines it, so a reader can trace every word and letter to its source.", "audience": [ "data-scientist", "auditor", "compliance", "product", "researcher" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Appendices_Vol1_Glossary_Notation_Index.pdf", "pages": 139, "bytes": 3201441, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Appendices_Vol1_Glossary_Notation_Index.pdf" } ] }, { "id": "A2", "kind": "appendix", "numeral": "A2", "label": "Appendix Volume 2", "title": "Metrics, Methods and Worked Examples", "docId": "VAI-COMP-APP2", "version": "1.0", "date": "2026-11-05", "chapters": null, "chapterCount": 0, "about": "The catalogs of fairness metrics (C) and explainability methods (D), the statistical methods (E) and worked numerical examples to recompute (H).", "keyMessage": "Every fairness number, explanation and statistical finding in the compendium can be traced to one exact definition, a stated method with its assumptions and a worked example that a reader can recompute before relying on it.", "audience": [ "data-scientist", "auditor" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Appendices_Vol2_Metrics_Methods_Examples.pdf", "pages": 158, "bytes": 3687822, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Appendices_Vol2_Metrics_Methods_Examples.pdf" } ] }, { "id": "A3", "kind": "appendix", "numeral": "A3", "label": "Appendix Volume 3", "title": "Law, Standards, Templates, Evidence and Bibliography", "docId": "VAI-COMP-APP3", "version": "1.0", "date": "2026-11-05", "chapters": null, "chapterCount": 0, "about": "The legal and regulatory matrix (F), standards (G), method worksheets and templates (I), the evidence taxonomy (J) and the bibliography of every source (K).", "keyMessage": "This volume collects in one place the law, the standards, the worksheets, the assurance vocabulary and the sources on which the compendium rests, each entry traced to the chapter that treats it.", "audience": [ "compliance", "auditor", "researcher" ], "files": [ { "label": "Full document", "file": "Validant_Compendium_Appendices_Vol3_Law_Standards_Templates_Bibliography.pdf", "pages": 218, "bytes": 4823016, "url": "https://validant.ai/compendium/pdf/Validant_Compendium_Appendices_Vol3_Law_Standards_Templates_Bibliography.pdf" } ] } ], "downloads": { "free": 5, "unit": "files per visitor", "then": "an email address at https://validant.ai/compendium unlocks every file", "excluded": "the front matter" }, "glossary": { "terms": 933, "source": "Appendix A, Controlled Glossary", "url": "https://validant.ai/compendium/glossary.json" } }