ReproStamp

ReproStamp is a web app (with optional CLI) that automatically captures the exact “recipe” behind a model training run: dataset fingerprints, code commit, environment packages, GPU/driver details, random seeds, config files, and key artifacts. It then generates a signed, shareable run card that teammates (or auditors) can replay to reproduce results or pinpoint why they can’t. This targets a persistent, expensive reality: many teams waste days chasing “it worked on my machine” training discrepancies, especially across multiple clouds and fast-moving experiments. ReproStamp integrates with common training workflows (PyTorch, Lightning, Hugging Face, Kubeflow, MLflow) and stores minimal metadata plus optional artifact pointers to your existing storage. It’s not a new training platform; it’s a thin reproducibility layer that sits on top of what you already use, reducing friction and political resistance to adoption.