Model Error Analysis Agent

One overall accuracy number hides where a model fails, so improvement work goes in the wrong direction. This agent takes a trained model and a labeled evaluation set, runs predictions and breaks errors down by segment, class and input feature to find where the model fails most. It groups similar errors and looks for patterns, such as poor results on a rare class or on short inputs. For each cluster it forms a hypothesis, like missing training examples, and checks it against the data. If the hypothesis does not hold, it tries the next one, including whether the labels themselves are wrong. It proposes a focused next step with the expected gain. You approve any change to the model or training data. Edge case: errors caused by mislabeled ground truth are flagged as a labeling issue.