diff --git a/assets/models/README.md b/assets/models/README.md
index 46de6f66..e9075390 100644
--- a/assets/models/README.md
+++ b/assets/models/README.md
@@ -18,9 +18,9 @@ cp ~/git/deep-darts-probe/models/dart_round<N>_withcal_saved_model/dart_round<N>
 Keep `kAutoScorerModelVersion` (in `dart_detector.dart`) in lock-step with the
 bundled stem — it is stamped onto every training capture's `model_version`.
 
-The current bundle is `dart_round24_withcal` (float32) — single net, classes
+The current bundle is `dart_round25_withcal` (float32) — single net, classes
 `{0:dart, 1:cal1, 2:cal2, 3:cal3, 4:cal4}`, imgsz 800, YOLO11n. It supersedes
-R21b; see the probe for R24's metrics (per-dart assist-mode segment accuracy on
+R24; see the probe for R25's metrics (per-dart assist-mode segment accuracy on
 the raw serve golden). The probe ranks rounds by **per-dart segment accuracy
 (assist-mode) on the raw serve golden**, not mAP/recall. Until a round is
 confirmed past the **88.9%** ship bar there is **no code-enforced emission
diff --git a/assets/models/dart_auto_scorer.tflite b/assets/models/dart_auto_scorer.tflite
index 8bc469c5..54585d25 100644
Binary files a/assets/models/dart_auto_scorer.tflite and b/assets/models/dart_auto_scorer.tflite differ
diff --git a/lib/features/auto_scorer/domain/detection/dart_detector.dart b/lib/features/auto_scorer/domain/detection/dart_detector.dart
index ccc1cc8a..3a6fdffe 100644
--- a/lib/features/auto_scorer/domain/detection/dart_detector.dart
+++ b/lib/features/auto_scorer/domain/detection/dart_detector.dart
@@ -12,7 +12,7 @@ const String kAutoScorerModelAsset = 'assets/models/dart_auto_scorer.tflite';
 /// every training capture so the probe can attribute each sample to the model
 /// that produced its predictions. Matches the probe's `pre_annotate` stem
 /// convention — bump this in lock-step with [kAutoScorerModelAsset].
-const String kAutoScorerModelVersion = 'dart_round24_withcal';
+const String kAutoScorerModelVersion = 'dart_round25_withcal';
 
 /// Runs the on-device detector and returns a [DetectionFrame] (cal points +
 /// dart candidates) for the tracker. Implementations preprocess the raw frame