This paper presents a visual analytics workflow for diagnosing binary classifiers using instance-level explanations. The workflow enables data scientists and domain experts to understand model decisions, identify weaknesses, and generate actionable improvement hypotheses through a structured multi-level exploration. Data Preparation and Model Training. The analyst prepares a feature set based on domain requirements, splits data into training and test subsets, and trains a binary classifier. Explanation Computation and Grouping. Prediction scores are computed for all test instances. Instance-level explanations are then generated for each data item by determining the minimal set of features whose removal changes the predicted label, treating the classifier as a black box. Instances sharing identical explanation feature sets are grouped into subsets representing distinct model decisions, and per-group statistics (prediction distributions, odds ratios with confidence intervals, discriminative feature rankings) are computed. Initial Overview. Three linked diagnostic views are rendered: prediction score histograms, confusion matrix, and ROC curve at the outcome level; a sortable explanation list with prediction distributions and statistical significance at the feature level; and an item-feature matrix with discriminative ordering at the instance level. The analyst reads the statistical summary to assess overall accuracy, error distribution, and generalization. Multi-Level Exploration. The analyst iteratively explores across the three levels — re-sorting explanations by frequency, significance, or uncertainty; filtering by prediction score range or feature search; and drilling down to instance-level feature matrices for specific explanation groups. Each cycle reveals decision patterns, significant or ambiguous predictors, misclassification sources, and feature associations that explain prediction errors. Exploration continues until sufficient patterns and root causes have been identified. Diagnostic Synthesis and Improvement. The analyst synthesizes diagnostic insights about model behavior, data limitations, and improvement opportunities. If the insights indicate that data engineering changes, additional features, or model adjustments could improve predictions, the analyst specifies those improvements and the cycle repeats with updated data and a retrained model. When the model is adequate or fundamental limitations are understood, the analyst formulates a final assessment of the classifier's practical utility.