Update ovarian_xenium/heldout_split/README.md

ovarian_xenium/heldout_split/README.md

@@ -109,20 +109,6 @@ Pearson r. The masks are fixed (seed 2026), so other models can be compared on t
 | 50 % of (cell, gene) entries masked | 0.165 | 0.167 |
 | no masking (what `get_model_output` returns) | 0.175 | 0.160 |
 
-As a reference, a truncated SVD of the same matrix at the same width (256) reaches median gene r =
-0.317. That is the ceiling for any linear encoder–decoder, and the model reaches about half of it.
-
-## Limitations
-
-* **One tissue section.** The test windows come from the same slide as the training windows, and
-  neighbouring windows are spatially autocorrelated. The scores measure generalisation to unseen
-  *regions of this section*, not to new donors or new slides.
-* **Cell masking** was chosen because it forces the model to predict a hidden cell from its
-  neighbours, and that neighbourhood dependence is what the attention read-out interprets. A
-  gene-masking model fills in missing genes of a visible cell better (r ≈ 0.25) but predicts a fully
-  hidden cell worse (r ≈ 0.13).
-* Attention-based net flow is a model read-out, not a measurement of ligand–receptor signalling.
-
 ## Citation
 
 Drummer, F., Jiménez, S. *et al.* InterScale. bioRxiv (2026). Code: