Every Percentage Point Counts: Production-Scale Visual Assessment of Scrap Quality for Low-Carbon Steelm

Recycling scrap through electric arc furnaces (EAFs) is the main near-term route to cutting the carbon footprint of steel, but its energy balance depends on a quantity that today is estimated by eye: the share of non-metallic contamination in each delivered railcar. Melting material that yields no steel wastes roughly 9--10\,kWh of grid electricity per tonne of charge for every percentage point of contamination, and the human scoring that governs both furnace planning and supplier payments carries an error that grows fourfold on the dirtiest deliveries. We build on ScrapSCORE, a benchmark created for this problem: 58{,}574 camera frames covering 2{,}636 railcar unloadings ($\approx$90{,}000\,t of scrap) at a working plant, each railcar scored independently by several inspectors, with fixed evaluation splits that test whether a model trained today still works after a year of seasonal and supply drift. Fine-tuned vision--language encoders reach inspector-level consistency on held-out railcars, yet every configuration we train loses a large share of its explained variance under temporal or cross-camera shift - and the model that transfers best is not the one that wins in-distribution. We quantify what measurement accuracy is worth in energy terms; the data, raw annotator scores and evaluation code are publicly available.