0031 nosamplecorrect mirrors numpys float64 upcast

0031 — NoSampleCorrect mirrors NumPy's float64 upcast, not requested-label accuracy

Status: accepted · Date: 2026-08-14

Context

sklearn.metrics.classification_report decides whether to print its support column as plain integers or as floats by asking multilabel_confusion_matrix whether any prediction anywhere in the dataset was correct — y_true[i] == y_pred[i] for some i — checked over every observed label, not only the labels a labels= argument requested. multilabel_confusion_matrix is pure Python — sklearn/metrics/_classification.py — and normally fills tp_sum/pred_sum/true_sum from _bincount, whose result is int64. When nothing at all matched, tp_bins is empty and the branch its own comment labels # Pathological case (sklearn/metrics/_classification.py:786, scikit-learn 1.9.0) seeds all three with xp.zeros(...) instead — a NumPy float64 array. NumPy then upcasts the whole assembled matrix to float64, so every support value downstream prints with a decimal point — 15.0, not 15 — even though nothing was weighted. The frozen oracle's all_wrong fixture pins exactly this: its reports field carries "15.0"/"17.0"/"18.0" supports rather than integers, for a 50-sample, unweighted, three-class target where nothing was ever right.

This is a different condition from Accuracy over the requested labels being zero. A sample whose true or predicted label falls outside the requested label set can still be the one correct prediction that makes the dataset-wide check true while Prf's per-class accuracy, computed only over the requested labels, is nonetheless zero — the two questions ("was anything, anywhere, right" vs. "was anything requested right") diverge exactly on that sample.

Decision

ConfusionMatrix.NoSampleCorrect is computed once, during Compute, over every observed label — not the requested subset — mirroring scikit-learn's own pre-restriction check rather than being derived later from the public, label-subset view. ReportText reads it to decide integer-vs-decimal support formatting, the same branch scikit-learn's dtype upcast produces.

Consequences

• ConfusionMatrix.NoSampleCorrect's <summary> carries a pointer here instead of restating the NumPy upcast mechanism.
• Verified by ReportTextTests.Support_stays_integral_when_a_correct_prediction_falls_outside_the_requested_labels for the "dataset-wide correct, requested-label accuracy zero" divergence this record's Context section describes, and by ReportTextTests.Renders_the_sklearn_table_character_for_character (tests/oracles/classification_metrics.json, fixture all_wrong) for the decimal-support formatting itself.