The classification scorecard, the parallel to the regression scorecard from an earlier episode. We build every metric up from the four cells of the confusion matrix (true positives, true negatives, false positives, false negatives), then show why accuracy lies on imbalanced data and how much of the metric zoo is just arithmetic on those four numbers. Covered: accuracy and the 99% trap, the DummyClassifier baseline, balanced accuracy, precision, recall, specificity, and the precision/recall tradeoff you steer with the decision threshold. F1 as the harmonic mean, F-beta for weighting recall against precision, and micro / macro / weighted averaging for multiclass. Then the probability threshold as the pivot: predict_proba versus decision_function, Youden's J, and scikit-learn 1.9's TunedThresholdClassifierCV and FixedThresholdClassifier. Curves: ROC (true positive rate versus false positive rate) and ROC AUC as the probability your model ranks a random positive above a random negative, equal to the normalized Mann-Whitney U. Why the precision-recall curve and average precision tell the truth on heavy imbalance when ROC AUC flatters a model that floods you with false alarms. Finally the probabilistic metrics: log loss, the Brier score, and calibration with reliability diagrams. History threads through: ROC from WWII radar operators and signal detection theory (Green and Swets 1966), precision/recall from information retrieval, van Rijsbergen's F-measure (1979), Youden's J (1950), Brier (1950), Hanley and McNeil (1982) on AUC and Mann-Whitney, and PR-over-ROC arguments from Davis and Goadrich (ICML 2006) and Saito and Rehmsmeier (PLOS ONE 2015). News brief up top: OpenAI's GPT-5.6 reaches general availability (Sol, Terra, Luna, Sol Ultra); xAI ships Grok 4.5; Google reportedly delays Gemini 3.5 Pro; NVIDIA drops open Nemotron models plus embedding datasets and training code; a self-evolving-agents research paper; and Norm AI and Monogram funding rounds.