[mahdi-shafiei]

Summary

-  "calibrax>=0.1.9",
+  "calibrax>=0.1.17",
-  # siblings; 0.1.6 adds substrax.runtime, substrax.artifacts and substrax.testing.
-  "substrax>=0.1.11",
+  # siblings; calibrax 0.1.17 requires 0.1.25.
+  "substrax>=0.1.25",

ExactAUROC delegates to calibrax's roc_auc. Before calibrax 0.1.17 a run of tied scores became several points of the ROC curve, so the area depended on the order of the tied records. 0.1.17 gives a run of ties one point, so a tied positive-negative pair counts half (the Mann-Whitney midrank convention). substrax moves to 0.1.25 because calibrax 0.1.17 requires it.

Evidence

• Before (calibrax 0.1.9), scores 0.5, 0.5, 0.2, 0.8:

  • test_a_tied_pair_counts_half_whatever_its_order returned 1.0 with labels 1, 0, 0, 1 and 0.75 with 0, 1, 0, 1.
  • test_every_score_tied_is_one_half returned 1.0.

  After (calibrax 0.1.17): 0.875 in both orders and 0.5, exactly.

• Locally on this branch, CPU: tests/losses 157 passed, tests/benchmarks 400, tests/utils 38, tests/pipelines 139, tests/operators 1,703, and the ownership test 1. pre-commit and uv lock --check pass. Every substrax name the tree imports resolves against 0.1.25.

Merge danger

Door: two-way.

Blast radius: four kinds of value DiffBio reports, each a calibrax fix since 0.1.9 (the CHANGELOG entry lists them):

DiffBio caller	calibrax change	effect
ExactAUROC	0.1.17, ties are one ROC point	tied scores count half a pair
clustering evaluation silhouette_score, calinski_harabasz_score	0.1.10, labels read as names	labels with gaps ({0, 2}) no longer count a cluster twice; match scikit-learn
batch-correction evaluation mmd	0.1.10, MMD_b (Gretton et al. 2012)	value changes for every input
masked_value_loss, MaskedValueLoss, compute_chromatin_guidance_loss (masked mse/mae)	0.1.15	a non-finite value at a masked-out position no longer gives NaN (or, for mae, a wrong value)

The softmax cross-entropy, entropy and unmasked mse call sites are unchanged. The example pages print silhouette and MMD values from 2026-03-19; whether they move is checked by rerunning examples/ecosystem/calibrax_metrics.py and scvi_benchmark.py before this leaves draft.