[ellispatrick]

Part of a stack of 8 pull requests from the October 2026 audit of ClassifyR (AUDIT.md on branch audit-2026-10). Each part is based on the one before, so they need to be merged in order. Each one can be accepted, changed or rejected on its own merits. Draft PR #125 holds all eight together. Every part installs and passes its own tests.

Stack, merge in order: #126 tests and core fixes → #127 interface fixes → #128 metrics and plots → #129 pathways and criss-cross → #130 maintenance → #131 faster cross-validation (identical results) → #132 shared folds (changes results) → #133 release notes. The regression harness and timings are in SydneyBioX/ProjectClassifyR (private).

Summary

Fixes to precision pathways, crissCrossValidate and the parameter classes.

Why and what

Precision pathways:

• Matching models: models are now matched to classifiers by assay name. They were matched by position, which only worked when the assays happened to sort alphabetically in the same order as the classifiers.
• Cost: a pathway's cost now counts every tier a sample passed through. A sample sent from clinical to miRNA was charged only for miRNA.
• Prediction applies the stored fold models directly, instead of re-running cross-validation on the test cohort and stratifying it by its labels. New samples no longer need classes, and Procedure 3's prediction went from about 19 min to 18 s.
• Cut-off: the confidence cut-off is a minimum (>=), as documented.
• Robustness:
  • the default mode works for clinical data plus one assay;
  • fixedAssays = NULL and list input work;
  • summary() matches weights by name;
  • assay names containing - work;
  • bubblePlot no longer silently drops pathways below 0.5 accuracy.

crissCrossValidate:

• Diagonal: the modelTest diagonal leaked test samples into feature selection. On noise it gave balanced accuracy 0.85, where honest CV gives 0.51. It is now each data set's own cross-validation, labelled as such.
• Features: data sets are restricted to their shared features.
• TOP is required only when runTOP = TRUE.

Classes:

• informative errors for one-step classifiers and unknown keywords;
• SelectParams("none") and plain functions work;
• TrainParams(balancing = ) is an error pointing to ModellingParams (it was silently ignored);
• show methods fixed;
• data documentation corrected.

Do results change?

Yes:

• Procedure 3 dollar figures change under cumulative costs. For example, clinical-miRNA goes from 8,750 to 8,900 end to end.
• Pathway confidences change, because prediction uses all fold models.
• The modelTest diagonal changes.

Evidence

289 tests pass (test-pathways.R, test-crisscross.R, test-classes.R).

Questions

• Prediction confidence averages all repeat × fold models (deterministic). The alternative is one random fold model per repeat.
• minAssaySamples at prediction makes one patient's routing depend on the rest of the batch.
• "Confidence" measures agreement between resampled models, not accuracy.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4

[DarioS]

Great speed improvements. Confidence is indeed repeated consistency - not accuracy.