Trust is assessed, not asserted.
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.
One book, a thumb tab for each of its twelve partsrest
Front matter · 38 pages · v1.1, 6 November 2026
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.
Written for every reader.
- F1Preface and Purpose
- F2Executive Overview
- F3Scope, Audience and Limitations
- F4Scientific and Editorial Method
- F5How to Use This Compendium
- F6Full Table of Contents
Hover a tab. Each one is a part, cut at the page where it begins, and the book opens there.
The work in figures
Chapters
Parts
Pages
Defined terms
Checked before publication
Reference entries
Claims
Formulas
Why this compendium
Three paragraphs, as published in version 1.1. Every chapter they cite is a handle: hover it to see what the chapter is and which printed part holds it.
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 (). 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 ( and ).
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 (), and that, when base rates differ between groups and prediction is imperfect, the main criteria cannot all hold at once (). Choosing among them is therefore unavoidable and normative (). A fairness metric computed without a stated question, a fixed analysis plan, validated measurements and quantified uncertainty is a number, not evidence (). And an assurance claim built on such evidence holds only for a stated scope, depth, criteria and date ().
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.
Find your way in.
By the task in front of you, by the question you need answered, by who you are, or by a term of the controlled glossary. The measure below is the whole work, one tick per chapter: your path lights up on it.
Thirteen foundational questions of fairness practice, each answered in full in the chapters named.
rest
Choose a task, a question, a reader or a term.
The reading path appears here, with the chapter to start with and the ones to continue with, and lights up on the measure above.
- 12Formalizing Algorithmic Decision-Making
- 13Probability and Statistical Foundations
- 14Independence: Demographic Parity and Related Criteria
- 15Separation: Equal Opportunity and Equalized Odds
- 16Sufficiency: Predictive Parity and Calibration
- 17Individual Fairness
- 18Causal and Counterfactual Fairness Criteria
- 19Impossibility Theorems
- 20Fairness Trade-offs
- 21Fairness Beyond Binary Classification: Scores, Regression and Ranking
- 22Contextual Selection of Fairness Criteria
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.
The same text, for agents and models.
This page is served as plain Markdown at the same address with an Accept header for text/markdown, and inside one index for language models. The licence travels with each copy.
- Index for language models/compendium/llms.txt
- This page as Markdown, or with Accept: text/markdown/compendium/index.md
- Structure as data: parts, chapters, reading paths, documents/compendium/index.json
- The controlled glossary, 933 terms with their chapters/compendium/glossary.json
- Terms for crawlers (Really Simple Licensing)/license.xml
- Every chapter as its own page and Markdown file, and the full text in one file: once the chapter text is exportedplanned
Licence. The AI Trust & Fairness Compendium, validant.ai (Glinz & Company GmbH), is published under Creative Commons Attribution-NonCommercial 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.