by-failure-diagnosis

You can diagnose failure modes by running Mann-Whitney U tests on PASS and FAIL design groups to identify statistically significant numeric feature differences. The tool applies Benjamini-Hochberg correction and reports p-values, q-values, and effect sizes to pinpoint discriminating metrics.

Statistical comparison of PASS versus FAIL groups requires Mann-Whitney U tests on numeric feature columns. This non-parametric approach detects distributional differences between successful and failed designs without assuming normal data distribution.

Yes, the diagnosis tool uses pandas and scipy to process your CSV data and calculate feature discrimination. It requires a PASS/FAIL status column alongside numeric feature columns to run the tests and generate actionable optimization recommendations.

Statistical failure post-mortem requires input CSV data containing a PASS/FAIL status column and numeric feature columns for canonical metrics like ipsae, iptm, plddt, rmsd, and cdr3_length. The scripts parse this tabular data to generate dashboards and exportable results.

You should shift to hypothesis debate when the diagnosis report shows no statistically significant discriminating features between PASS and FAIL groups. If Benjamini-Hochberg corrected q-values indicate no meaningful feature separation, simple parameter changes and threshold tuning will likely be insufficient.

Yes, Benjamini-Hochberg correction is applied during failure diagnosis to control the false discovery rate across multiple Mann-Whitney U tests. It adjusts p-values into q-values, ensuring reported discriminating features between PASS and FAIL designs are statistically reliable.