diff --git a/antipasto/train/inner_contrastive_loss.py b/antipasto/train/inner_contrastive_loss.py
index 5cd500b..6e2b23b 100644
--- a/antipasto/train/inner_contrastive_loss.py
+++ b/antipasto/train/inner_contrastive_loss.py
@@ -372,16 +372,19 @@ def contrastive_steering_loss_with_ref(
 
     # Concentration-aware: weight by subspace focus (||proj|| / ||full||)
     # focus ∈ [0, 1] = fraction of delta energy in loss subspace
-    # Clamped to [0, 1] because numerically can exceed 1 due to Fisher weighting mismatch
+    # Compute in RAW space (undo Fisher) so focus is apples-to-apples with delta_full
     cos_pos_ref_used = cos_pos_ref
     cos_neg_ref_used = cos_neg_ref
     focus_pos = None
     focus_neg = None
     if delta_pos_norm_full is not None and delta_neg_norm_full is not None:
-        proj_norm_pos = antisym_pos_agg.norm(dim=-1)  # [b]
-        proj_norm_neg = antisym_neg_agg.norm(dim=-1)  # [b]
-        focus_pos = (proj_norm_pos / delta_pos_norm_full.clamp(min=eps)).clamp(max=1.0)
-        focus_neg = (proj_norm_neg / delta_neg_norm_full.clamp(min=eps)).clamp(max=1.0)
+        # Undo Fisher normalization to get raw-space norms for concentration
+        proj_raw_pos = antisym_pos_agg * std_for_div.unsqueeze(0)  # [b, r]
+        proj_raw_neg = antisym_neg_agg * std_for_div.unsqueeze(0)  # [b, r]
+        proj_norm_pos = proj_raw_pos.norm(dim=-1)  # [b]
+        proj_norm_neg = proj_raw_neg.norm(dim=-1)  # [b]
+        focus_pos = proj_norm_pos / delta_pos_norm_full.clamp(min=eps)
+        focus_neg = proj_norm_neg / delta_neg_norm_full.clamp(min=eps)
         if focus_softness > 0:
             focus_pos = focus_pos.pow(1.0 - focus_softness)
             focus_neg = focus_neg.pow(1.0 - focus_softness)