Case study 01

Tabular Uncertainty

Rebuilding TabM to test disagreement as a measurement.

A from-scratch PyTorch reproduction asking whether implicit ensemble disagreement can flag distribution shift without giving up TabM's efficiency.

Evidence at a glance

Question: Can implicit ensemble disagreement flag distribution shift?
Method: Eight-member disagreement illustration with a multi-seed benchmark plan.
Caveat: Illustration only; not a trained TabM checkpoint.

Can disagreement expose distribution shift?

The reproduction implements TabM and TabM-mini from scratch, then treats the implicit ensemble as a possible measurement surface rather than only a way to improve predictions.

What is being built

A PyTorch reproduction with regression, binary, and multiclass evaluation, multi-seed runs, member pruning, calibrated disagreement, and automated logs.

What remains unproven

The interaction below demonstrates the proposed reading. Its values are illustrative and do not come from a trained TabM checkpoint.

Inspect the ensemble

Change the specimen, then include or exclude members to see how consensus and spread produce the qualitative read.

Illustration of the disagreement test, not a trained TabM checkpoint.