evidence/claim5/limitations.md

# Claim 5 limitations and deviations

This is a falsification of the exact “consistent improvements” quantifier,
using the paper's complete reported result domain and a different-seed
corroborating table. It does not rerun the authors' single-H100 image training
on CPU, estimate per-seed significance from unavailable raw seed values, or
deny the selected-step improvements for KL and chi-squared.

Because the imported Claim 5 is conjunctive, one valid counterexample to the
consistency conjunct falsifies that exact claim. The broader statement that
JKO can help some objectives remains supported by the same table.