Versioned Model Cards and Dataset Cards
Transparency is not an attached PDF: it is a versioned card with status.
Framework / standard: EU AI Act Art. 11 (Annex IV) · Art. 10 · ISO/IEC 42001 A.7.5
The limitations of your AI should not be scattered prose in a wiki. Article 11 of the EU AI Act (and its Annex IV) requires technical documentation for high-risk systems; ISO 42001 A.7.5 requires documenting data resources. Linedat materialises this as versioned cards.
The Model Card documents the intended use, out-of-scope use, intended users, performance, limitations and known biases of each model. The Dataset Card documents the grounding corpus of your RAG.
Model Card per AI system
Each Model Card is linked to an AI system and includes intended use, out-of-scope use, performance metrics, limitations and known biases. It carries a status (draft / published / needs update) and obsolescence flags, so an outdated card does not go unnoticed. It is the answer to Art. 11 without relying on a document that nobody maintains.
Dataset Card for the RAG corpus
The Dataset Card documents the corpus that feeds your RAG: motivation, composition, pre-processing (including sanitisation) and declared coverage biases. Quality is derived from the catalogue (from AssetQualityScore), not duplicated. A RAG system without a Dataset Card generates a finding — the corpus stops being a black box.
The limits (what we do not claim)
Cards are structured and versioned documentation, not the result of an automated model evaluation: the responsible person fills in the metrics and biases. The documented provider is a declarative snapshot that may fall out of sync with the model actually running if not updated.
How Linedat helps
Linedat turns the transparency required by Art. 11 and A.7.5 into versioned cards with status and obsolescence warnings — recognisable to an auditor, linked to the actual system and corpus.
Related capabilities
The AI Act audits systems, not API calls. Every AI system inventoried with its risk class.
FRIA: Fundamental Rights Impact AssessmentThe FRIA required by EU AI Act Art. 27, versioned and linked to the AI system — not a Word document sitting in a folder.
Executable Data Quality (Real SQL, 6 Dimensions)Quality stops being an opinion: rules that execute real SQL against the source, in 8 dialects.