The decisive point is failure isolation: the combination of Options A and C keeps the agent's decision path observable instead of burying behavior inside one prompt or one service. Together, A states "Design ablation studies systematically varying individual parameters while holding others constant to isolate each parameter's impact on agent behavior and performance."; C states "Implement A/B testing frameworks comparing temperature, top-k, and top-p variations while measuring task-specific quality metrics and user satisfaction scores.", so the answer covers both sides of the requirement instead of solving only the model or only the infrastructure layer. Ablation isolates parameter impact; A/B testing validates it against user-facing quality.
Random simultaneous changes destroy causal interpretation. The implementation detail that matters is repeatable benchmark suites that separate accuracy, cost, latency, reliability, and human satisfaction rather than blending them into one vague score. The stack-level anchor is clear: the NVIDIA stack makes it possible to correlate model-serving metrics with workflow events and user-visible task failures. The losing choices mostly optimize for short-term convenience; offline benchmarks alone cannot expose live API failures, schema drift, queue saturation, or feedback-driven dissatisfaction. That is the difference between an agent that works in a notebook and an agent that remains reliable in production.