An open-source tool that answers one question before you spend a GPU-day: is this labelled object-detection dataset safe to train on, and what has to be fixed first? It is built for engineers handed a dataset they did not create, such as a vendor delivery, a Roboflow export or a merge of internal collections.

It catches the defects that never show up by eyeballing images: train/val leakage that inflates mAP, zero-area boxes that put NaNs in the loss, pixel coordinates left unnormalised, sub-pixel boxes the dataloader silently drops, and classes too sparse to measure. It returns a verdict (train or blocked), a triage report and an ordered fix plan that it can apply.

Every finding comes from deterministic detectors, not a language model. A local LLM (Qwen) only orchestrates the audit and ranks what matters. scan runs offline in seconds, audit produces the report, fix plan and full trajectory, and the project ships with 218 offline tests under Apache-2.0.