Match each ops discipline to the artifact that primarily moves through its pipeline

Each ops discipline owns one primary deployable artifact, even though the others orbit it as supporting concerns. **DevOps** ships application code. The artifact is a commit that becomes a container image that gets deployed via CI/CD to a runtime environment. The gate is unit tests, integration tests, and code review. **MLOps** ships trained models. The artifact is a model weight file plus the dataset and training code that produced it. The gate is offline evaluation on a held-out set. The pipeline includes data preparation, training runs, experiment tracking, model registry, and serving. **LLMOps** ships prompts, retrieval indexes, tool definitions, routing rules, and rarely a fine-tune. The base model is usually a third-party black box (Claude, GPT, Gemini) or a pinned open-weights snapshot; the team iterates in the layers above the model. The gate is eval as CI golden-set replay rather than held-out accuracy. **DataOps** ships ETL pipeline DAGs and data-quality contracts. The artifact is the transformation logic that moves data between systems and the contracts that gate downstream consumers from broken upstream data. **PromptOps** is the youngest of the five. It ships prompt templates managed in a registry (Langfuse, Humanloop, Pezzo) independently of code deploys. It overlaps with LLMOps but is narrower: PromptOps is specifically about the prompt as versioned artifact discipline. The disciplines share infrastructure (Kubernetes, Git, observability) but the primary artifact and deploy gate differ.

Each ops discipline is defined by its primary deployable artifact and the gate that artifact has to pass before reaching production. Five disciplines, five artifacts. **DevOps** ships application code. Commit becomes container image becomes deployment. Gate: tests and review. **MLOps** ships trained models plus the data and training code that produced them. The model is the artifact; the data and training are the lineage. Gate: offline evaluation on a held-out set, with experiment tracking (MLflow, Weights and Biases, Comet) recording every run. **LLMOps** ships the layers above the model. Prompt diffs, retrieval indexes, tool definitions, routing rules, and very occasionally a fine-tune. The base model is a third-party black box or a pinned open-weights snapshot; iteration almost never involves retraining. Gate: eval as CI golden-set replay, with a typical pipeline using shadow mode, dogfood, and graduated canary. **DataOps** ships ETL pipeline DAGs (Airflow, Dagster, Prefect) and data-quality contracts on tables (Great Expectations, Soda, dbt tests). The artifact is the transformation logic and the contract gate that protects downstream consumers. Often the upstream supplier for MLOps and LLMOps. **PromptOps** ships prompt templates managed in a registry, versioned, audited, and deployable independently of code. It is a sub-discipline of LLMOps that emerged because prompt iteration happens on a different cadence than code deploys and benefits from its own audit and rollout surfaces. **The throughline.** Each discipline has one primary artifact even though the others orbit it as supporting concerns. Collapsing them into one operations team works at startup scale but fragments at production scale because the deploy gates, audit needs, and on-call surfaces differ.