AI Act Article 17 technical documentation: what level of model architecture detail do auditors actually require?

We're preparing for our first EU AI Act readiness audit and hitting a practical wall on Article 17 (technical documentation). The regulation says providers must draw up technical documentation "before placing a high-risk AI system on the market" and it must contain "detailed information about the system's architecture, training data, and performance metrics." In practice, how granular are your auditors expecting you to go? - Full layer-by-layer architecture diagrams? - Training data lineage for each feature (not just the final dataset)? - Hyperparameter tuning logs and ablation study results? - Model card equivalents with bias/fairness metrics per demographic slice? Jurisdiction: EU, DE Confidentiality Acknowledged: true For context: we're a mid-size SaaS with 3 ML models in production (fraud detection, churn prediction, document classification). None are customer-facing AI, but two process EU personal data at scale. Our external counsel says "document everything," but our engineering lead says that's a 6-month project. Looking for what actually moved the needle in a real audit.