How did your team operationalize Art. 22 GDPR automated decision disclosures for ML model retraining cycles?

We're running ML-driven fraud scoring that triggers account holds — clearly falling under Art. 22 GDPR scope for automated decision-making with legal/similar effects. The challenge is that model retraining (monthly) changes decision logic, so the 'meaningful information about the logic involved' that Art. 22(3) requires becomes outdated between disclosures. Questions for peers: 1. Do you maintain a model-card-style disclosure document that updates with each retrain? 2. How do you handle the 'right to obtain human intervention' when the model retrains — do you notify affected data subjects proactively? 3. What's your documentation approach for feature importance shifts (e.g., SHAP value changes) that affect decision rationale? We've looked at CNIL guidance and the EDPB guidelines, but the operational detail is sparse. Particularly interested in how fintech or insurance teams handle this. Jurisdiction: DE, EU Framework: GDPR Art. 22, EDPB Guidelines 3/2019 on automated individual decision-making.