reproducibility

What problem does it solve? Experimental results often cannot be re-derived because seeds, configs, environments, and logs were never recorded at run time, making every number untrustworthy and every conclusion unverifiable. ## Core Features & Use Cases - Day-one record template: Creates an experiment directory with config.yaml, seeds.txt, env.lock, run.sh, logs/, and outputs/ before any run starts. - Calibration checkpoint: Reproduces a known published result before trusting any new measurement, validating the pipeline as an instrument. - Sealed test set: Splits and hash-archives the test set before preprocessing so it participates in no computation until final evaluation. - Full trace archiving: Stores every run's command, stdout/stderr, random seed, and git commit together with results. - Use Case: Before launching a new model training run, use this Skill to set up the experiment directory, lock the environment, rerun the baseline's published numbers as a calibration check, and seal the test set under a hash. ## Quick Start Set up a reproducible experiment directory with seeds, locked environment, and a calibration checkpoint before I run my first training job.