Production AI systems are socio-technical systems whose behavior emerges from interacting states across data, code, execution environments, model or provider state, prompts, retrieval context, tools, policies, evaluators, infrastructure, and human intervention. This paper develops an evidence-centric framework that treats Machine Learning Operations (MLOps) and Large Language Model Operations (LLMOps) as disciplines for controlled change in such systems. The framework integrates evidence from recent MLOps architecture studies, LLMOps research, generative-AI evaluation literature, current platform documentation, security guidance, and AI governance frameworks. Five contributions are developed: (1) a multidimensional architecture lens that extends infrastructure-friction versus operational-yield analysis into an explicit trade-off space; (2) executable provenance, which defines reproducibility as reconstruction of materially causal execution state; (3) a composable MLOps/LLMOps architecture separating evidence, evaluation, orchestration, compute, knowledge, serving, telemetry, policy, and human control; (4) evaluation as a release-control mechanism in which the evaluator is itself versioned; and (5) a capability-coverage model for organizational readiness. The paper derives eight falsifiable hypotheses and specifies a controlled benchmark protocol spanning supervised ML, parameter-efficient fine-tuning, retrieval-augmented generation (RAG), and tool-using agents. An author-supplied Kubeflow-oriented repository is examined as an implementation artifact and is explicitly distinguished from evidence about current Kubeflow Pipelines v2 behavior. The resulting framework provides measurable constructs for studying architecture, provenance, evaluation, security, governance, organizational capability, and production economics without assuming a universal platform configuration.