Upload eval/swe_v10_results.json with huggingface_hub

eval/swe_v10_results.json

@@ -0,0 +1,38 @@
+{
+  "v10_direct": {
+    "success": 0.766,
+    "avg_cost": 0.1878,
+    "costRed": 40.7
+  },
+  "v10_cascade": {
+    "success": 0.756,
+    "avg_cost": 0.1767,
+    "costRed": 44.2
+  },
+  "v10_feedback": {
+    "success": 0.848,
+    "avg_cost": 0.2014,
+    "costRed": 36.4
+  },
+  "v8_synthetic": {
+    "success": 0.658,
+    "avg_cost": 0.3534,
+    "costRed": -11.6
+  },
+  "frontier": {
+    "success": 0.782,
+    "avg_cost": 0.3167,
+    "costRed": 0.0
+  },
+  "oracle": {
+    "success": 0.87,
+    "avg_cost": 0.0624,
+    "costRed": 80.3
+  },
+  "always_cheap": {
+    "success": 0.632,
+    "avg_cost": 0.0142,
+    "costRed": 95.5
+  },
+  "key_finding": "v10 trained on REAL SWE-Router data achieves 36-44% cost reduction vs 8% for synthetic-trained v8. v10_feedback achieves HIGHER success than always-frontier (84.8% vs 78.2%) at 36.4% cost reduction."
+}