AI model documentation completeness checker

Intended purpose and use context

You have written down what the system or model is for, who is meant to use it, in what context, and the conditions, assumptions, and limits of its intended use, including reasonably foreseeable misuse to avoid. (EU AI Act Annex IV(1); Annex XI for GPAI models.)

General description, versions, and acceptable use

A reviewer can follow what the system does and how it is provided (as a product, an API, or an embedded component), which version or release this documentation covers, and any acceptable-use terms or licence attached to it. (Annex IV(1); Annex XI model description.)

Data sources and provenance

You record where the training, validation, and test data came from, how it was collected or obtained, and the licensing or legal basis under which you use it. (Annex IV(2); Article 10; Annex XI data information.)

Dataset composition and preparation

You document the datasets used to build and evaluate the system: their size and main characteristics, how they were labelled, and the cleaning, filtering, or other preparation applied. (Annex IV(2); Annex XI data curation.)

Data quality and bias examination

You document how data quality and relevance are judged, and how possible biases in the data and their effects on the people the system affects are examined and addressed. (Article 10; Annex XI bias-detection measures.)

Design choices and architecture

You document the key design decisions, the general logic and method of the system, the model type and main parameters at an appropriate level, and the rationale for the principal choices, without having to disclose trade secrets. (Annex IV(2); Annex XI architecture and parameters.)

Performance metrics and evaluation

You report the metrics used to measure accuracy and performance, why those metrics fit this system, the results on representative data, and how the system was tested and validated. (Annex IV(2), (3), and (4); Annex XI evaluation results.)

Known limitations and expected accuracy

You state the system's known limitations, the conditions under which performance degrades, the level of accuracy users should expect, and any group for which it may perform differently. (Annex IV(3).)

Risks to people and mitigations

You document the foreseeable risks the system poses to health, safety, fundamental rights, or other people, and the measures taken to identify, evaluate, and reduce them. (Annex IV(5); Article 9.)

Human oversight measures

You document how people can understand, monitor, intervene in, or override the system, and the technical and organisational measures built in to make that oversight effective. (Annex IV(2) and (3); Article 14.)

Information and instructions for users

You give the people deploying or using the system clear instructions: its capabilities and limitations, the conditions required for safe use, and what is expected of a human overseeing it. (Article 13.)

Robustness, accuracy, and cybersecurity

You document the measures that keep the system accurate and resilient against errors and adversarial conditions, and that protect the system and its data against cybersecurity threats, consistent with its intended use. (Annex IV(2); Article 15.)

Versioning, change, and post-deployment monitoring

You version the documentation with the system, record the relevant changes made across its lifecycle, and have a plan to monitor how it performs once it is in use. (Annex IV(6) and (9); Article 72; ISO/IEC 42001:2023 lifecycle expectations.)

Third-party components, standards, and resources

You identify the third-party tools, pre-trained or general-purpose models, and other externally-sourced components the system relies on, the standards or alternative solutions you applied, and, for larger models, the computational resources used. (Annex IV(2) and (7); Annex XI computational resources.)