Add dalpha-llm-router submission

model_cost/model_cost.json

@@ -386,5 +386,17 @@
   "google/gemma-4-31b-it": {
     "input_token_price_per_million": 0.08,
     "output_token_price_per_million": 0.35
+  },
+  "gemini-3.8-flash": {
+    "input_token_price_per_million": 0.75,
+    "output_token_price_per_million": 3.75
+  },
+  "gpt-6-luna": {
+    "input_token_price_per_million": 0.1,
+    "output_token_price_per_million": 0.5
+  },
+  "anthropic/claude-haiku-5.5": {
+    "input_token_price_per_million": 0.1,
+    "output_token_price_per_million": 0.5
   }
 }

router_inference/config/dalpha-llm-router.json

@@ -0,0 +1,13 @@
+{
+  "pipeline_params": {
+    "router_name": "dalpha-llm-router",
+    "router_cls_name": "external_prediction_file",
+    "models": [
+      "gemini-3.8-flash",
+      "google/gemma-4-31b-it",
+      "anthropic/claude-haiku-5.5",
+      "deepseek/deepseek-v4-pro"
+    ],
+    "description": "Dalpha LLM Router -- routes each query across a 4-model pool (gemini-3.8-flash at thinking_level=low, gemma-4-31b-it, claude-haiku-5.5, deepseek-v4-pro with thinking off) using a zero-shot query classifier and a per-task routing table calibrated on external data only (public source benchmarks, items disjoint from RouterArena). No RouterArena data was used to build or tune the router."
+  }
+}