reproducibility-standardslisted

## Purpose Make every result reproducible from source control + logged state alone. ## When to use it Setting up a new project, starting a new experiment, or preparing work for handoff. ## Checklist - **Seed all random generators**: numpy, `random`, torch, tensorflow, and any other stochastic components. Log seeds. - **Pin dependencies**: use lockfiles (`requirements.txt`, `poetry.lock`, `conda-lock`) — not loose version ranges. - **Capture environment**: Dockerfile or conda env.yml that reproduces the exact runtime. - **Version data**: use DVC, hash-based manifests, or snapshot the exact dataset version used. - **Log transforms**: every preprocessing step (scaling, encoding, imputation) must be logged and reproducible. - **Pipeline DAG**: capture the full pipeline graph with input/output hashes per step. ## Expected output A project setup where `git clone → install → run` reproduces the same results, verified by comparing logged metrics.