You already know: evaluation per slice and confidence intervals (D10, D2); data provenance, PII scrubbing and datasheets (D16); red-teaming and agent controls (this domain); calibration (D3).
Why this matters for Atlas. "Responsible AI" sounds like a values statement, but in engineering it is a set of measurable properties and processes: error rates compared across groups, personal data handled with explicit limits, documentation that says what a model is for and where it fails, legal risk classification that decides which controls are mandatory, and humans with real authority over consequential decisions. This lesson treats each as an engineering problem: fairness metrics and their trade-offs, counterfactual bias testing for LLMs, privacy techniques, model and system documentation, and the regulatory frame (EU AI Act, NIST AI RMF) translated into release requirements.
1. Fairness: which errors, for whom
The question. What does it mean for the at-risk predictor to be "fair"?
Start from nothing. Fairness metrics compare a model's behaviour across groups (e.g., students who work part-time vs those who don't, or by language background):