Conditional Transport Lab
Explore numeric correction while keeping categorical facts and generated labels fixed.
Runs in your browser. No account, GPU or model API required.
Overview
This anonymous computational artifact demonstrates label-only empirical rank mapping and a training-fitted categorical conditional transport. It preserves generated labels, categorical values and row count while changing numerical assignments. A saved-result explorer separates generator and fitting-seed summaries. A fictional training table makes stable ties, mixture weights and functional-dependency anchors inspectable. A small released-benchmark prediction pair permits direct AUROC and Brier checks.
Fit on training data. Compare paired outputs.
For each label and numerical field, a training-only tree partitions the categorical predictors. A leaf reference mixes its empirical distribution with the label-wide distribution. Within-leaf native-score ranks select values from that mixture.
λ_leaf = training leaf rows / (training leaf rows + 64) F_mix = λ_leaf F_leaf + (1 − λ_leaf) F_label ΔAUROC = conditional transport − label-rank
The mapping exercise uses fictional data and a frozen, training-fitted tree. It does not retrain TabDiff or TabSyn. Saved results and fictional mappings remain separate. There is no medical recommendation or privacy guarantee.
Inspect generator-specific effects.
Fitting-seed means and draws are not independent datasets. Archived intervals retain their original fixed-fit or fit/group interpretation. Changing the filter does not create a new confirmatory test.
Fictional quantile-mapping lab
The 192 training rows and 24 generated rows are invented. Labels, categories and row count stay fixed. A constant-field anchor remains visible in the downloaded data.
Recompute a complete saved prediction pair
Heart Disease · TabDiff fit 17 · draw 1701 · CatBoost · 184 held observations. This one pair is a numerical example, not the pooled extension result.
Labels: Janosi, Steinbrunn, Pfisterer & Detrano, Heart Disease, UCI, DOI 10.24432/C52P4X, CC BY 4.0. Study outputs supply the saved probabilities. No identifiers or clinical feature rows are included.
From the page to the code.
The download includes this website, numerical samples, research reference code, provenance and executable checks. The examples are usable without access to a hospital or private account.
unzip transport-lab.zip cd research-artifact python reproduce.py # Browse the static website locally: python -m http.server 8766 --directory dist
Download the complete installation package · Inspect provenance
Saved results are computational evidence. This artifact does not establish independent clinical validation, medical suitability or acceptance of a paper.