experiment-trackinglisted

## Purpose Ensure every experiment is reproducible from its logged state alone. ## When to use it Every training run, every data transformation, every evaluation. ## Checklist - **Run naming**: `{date}-{objective}-{attempt}` (e.g. `2026-08-12-classifier-lr-search-03`). - **Seed everything**: numpy, python `random`, torch, tensorflow — log which seeds were used. - **Log parameters**: hyperparameters, data splits, preprocessing choices, model architecture. - **Log metrics**: final metrics per split (train/val/test), per-epoch metrics if relevant. - **Log artifacts**: model weights, predictions, feature importance plots, confusion matrices. - **Log environment**: Python version, dependency versions (lockfile or `pip freeze`), git commit hash. - **Compare runs**: use the tracker's comparison view or export to a structured format. - **Recover**: from a logged run, you should be able to reproduce the exact result. ## Expected output A tracked run that another person or agent can reproduce without asking the original author for details.