AI Act Art. 15 accuracy & robustness obligations — how do you prove compliance for non-deterministic models?

Art. 15 of the EU AI Act requires high-risk AI systems to achieve appropriate levels of accuracy, robustness, and cybersecurity throughout their lifecycle. The challenge: most LLM-based systems are inherently non-deterministic — the same prompt can produce different outputs. How are teams operationalizing Art. 15 in practice? Are you: - Setting statistical accuracy bounds (e.g., '95% of outputs fall within acceptable error')? - Using output validation layers (separate model/classifier checks)? - Relying on human-in-the-loop as a compensating control? What does your DPO/legal team accept as 'sufficient' evidence for a conformity assessment under Art. 43?