"The model card is the receipt your auditor asks for."
Compass, Documentation-Disciplined AI Agent
Chapter 54 marked AI-generated outputs; this chapter documents the systems that produced them. Model cards, datasheets for datasets, system cards, audit trails, and the explainability disclosures that high-stakes decisions now require.
The disclosure side of trustworthy AI: model cards, datasheets for datasets, system cards for frontier deployments, audit trails, and explainability for high-stakes decisions. This chapter is the second half of the provenance and transparency split: the first half (Watermarking and Provenance) covered how to mark and trace AI-generated content; this half covers the structured documentation and explainability artifacts that let consumers, regulators, and auditors understand what a model is, what it was trained on, and why it produced a given output.
Chapter Overview
Transparency is the documentation layer of LLM governance: model cards that travel with a checkpoint, datasheets that travel with a corpus, system cards for deployed frontier systems, audit trails for compliance, and explainability for high-stakes decisions. This chapter teaches the canonical documentation primitives, with real examples from Llama-3 and Claude 3.5 Sonnet, and shows how each artifact maps to procurement, compliance, and regulatory workflows.
Documentation has become regulatory infrastructure. By the end of this chapter you will know what to write, who reads it, and how each artifact maps to a specific compliance or accountability question.
- Architect a model card for an LLM release that satisfies EU AI Act transparency duties.
- Apply the datasheets-for-datasets template to a training or evaluation corpus.
- Compose a system card for a deployed frontier system, beyond per-model documentation.
- Design tamper-evident audit-trail logging that supports regulatory review and post-incident analysis.
- Build explainability outputs that satisfy regulatory disclosure for high-stakes decisions.
Prerequisites
- Regulation and compliance from Chapter 53
- Watermarking and provenance from Chapter 54
- Pretraining and data curation from Chapter 6
Sections
- 54.6 Model Cards: Anatomy, Examples, Use in Procurement The structured documentation that travels with a trained model, with examples from Llama-3 and Claude 3.5 Sonnet. Intermediate
- 54.7 Datasheets for Datasets Per-dataset documentation that travels with training and evaluation corpora. Intermediate
- 54.8 System Cards and Frontier System Disclosures System-level disclosures for deployed frontier AI, beyond per-model cards. Intermediate
- 54.9 Audit Trails and Logging for Compliance Tamper-evident logging that supports regulatory audit and post-incident review. Advanced
- 54.10 Explainability for High-Stakes Decisions Explanations that satisfy regulatory disclosure requirements and support human oversight in consequential domains. Advanced
What's Next?
This chapter begins with Section 54.6: Model Cards: Anatomy, Examples, Use in Procurement. Each section walks one transparency artifact from anatomy to procurement workflow, so we recommend reading them in order.