model card: describe v4 only, drop the v1/v2 comparison

README.md

@@ -16,14 +16,9 @@ pipeline_tag: image-classification
 # RootScope: Cross-species Root Cell-Type Classification from Confocal Microscopy Images
 
 Trained weights for **[RootScope](https://github.com/ct-tranchau/Rootscope)**.
-This repo holds every published version; the package picks one with
-`--model-version` (default **v4**).
-
-| version | folder | what it is |
-|---|---|---|
-| **v4** | `v4/` | the manuscript model. 915 features, LightGBM per refinement round, 3 seeds, radial-order prior |
-| v2 | repo root | 864 features, RandomForest + XGBoost + LightGBM ensemble, 10 rounds |
-| v1 | `v1/` | 480 features, DINOv2 ViT-S/14 |
+The manuscript model (v4) lives in `v4/`: 915 features, LightGBM per
+refinement round, 3 seeds, radial-order prior. The package downloads it on
+first use.
 
 ## v4 files
 
@@ -53,7 +48,7 @@ On species absent from training, expect roughly 0.70.
 
 ## How v4 predicts
 
-1. Cellpose-SAM segments the TIFF (zero-shot, unchanged from v1/v2).
+1. Cellpose-SAM segments the TIFF (zero-shot).
 2. 96 morpho-topological features per cell + 768-d embedding from the
    fine-tuned backbone.
 3. Round 1: classify with the neighbor / layer context set to "unknown".