Transit plans a blue/green cutover of a safety-related GenAI assist model but has no automated evaluation suite comparing safety metrics. What MLSecOps step should precede cutover?
Select an answer to reveal the explanation.
Short Explanation
Blue/green without a pre-flight scorecard is changing runways in fog. Evaluate safety metrics automatically before you flip traffic. After-the-fact complaints, killing metrics, and removing rollback color are not gates.
Full Explanation
Transit planning a blue/green cutover of a safety-related GenAI assist model without an automated evaluation suite comparing safety metrics should run automated evaluation of safety and quality metrics before any traffic switch. Blue/green provides capacity to flip quickly, but without a pre-flight scorecard you may promote regressions in refusal behavior, hallucinated guidance, or other safety-adjacent outputs that affect riders and operators. Automated comparison against the live blue baseline makes unsafe deltas visible before green receives production traffic.
Cutting over first and measuring metrics only after resident complaints fails because it uses the public as the test suite. Disabling safety metrics to avoid blocking releases fails because it prioritizes ship speed over transit safety signals. Deleting the blue environment so green is the only option fails because it removes rollback capacity that blue/green is meant to preserve.
Exam caveat: infrastructure patterns (blue/green) do not replace evaluation gates for safety-related civic GenAI; exams expect metrics-before-switch.
Operational check: define safety/quality thresholds; run the suite on green against held-out and shadow samples; block DNS/traffic flip on failed gates; keep blue warm until green proves stable; rehearse rollback using the evaluation dashboard as the go/no-go record.