Compared origin/feat/isup-inference with origin/merge-unify-vits-into-cleanup-vits without editing ve2e.
Traced the vxData standard pipeline, patient filtering, study selection, training-table output, data module, clinical dataset preprocessing, unified model forward, checkpoint loading, and classification metrics.
Reported exact reusable APIs and line references to the parent agent.
Identified unsafe split relabeling and label-driven study selection as evaluation integrity risks.
Identified obsolete feature commits and the inference package path collision caused by the target repository restructure.
Lessons Learned: Pitfalls
The knowledge repository does not contain wiki/workstreams/clinical/index.md or its SOP index, although the mandatory begin-work instructions say to read them. The missing workstream entry should be created or the instructions should describe the fallback.
Similar ISUP audit recaps already existed in the inbox. A quick inbox search before creating another recap would reduce duplication.
Lessons Learned: Improvements
A short durable ve2e note describing the canonical vxData split names (train, val_nb, val_rp, benchmark) and which pipelines may relabel them would make future inference audits faster.
Document that select_primary_study uses mask origin and ISUP, so it is unsuitable for leakage-free evaluation study selection unless the cohort definition explicitly requires that policy.
