Reproducible Experiment Records

Connect code, data, and results

Tag each run with its Git commit (the exact source-code snapshot) and Data Version Control (DVC) data revision. Save the prompt template and evaluation output as artifacts. Never record passwords, API keys, or raw prompts that contain sensitive data.

Use MLflow's run context - the with mlflow.start_run(): block - so an exception still closes the run and marks it failed. Give experiments stable names and runs useful descriptions instead of relying only on generated IDs.

Goal: The MLflow lab contains a script pointed at the wrong store. Correct the URI and prove the required run exists in the shared local backend.