Add ainative-router submission (AINative Router, content-only routing)

llm_inference/model_inference.py

@@ -191,6 +191,10 @@ def _get_provider(self, model_name: str) -> str:
             "xiaomi/mimo-v2-flash:free": "openrouter",
             "openai/gpt-oss-120b": "openrouter",
             "qwen/qwen3-235b-a22b-2507": "openrouter",
+            "openai/gpt-6-luna": "openrouter",
+            "deepseek/deepseek-v4.1-flash": "openrouter",
+            "deepseek/deepseek-v4-flash-0731": "openrouter",
+            "google/gemma-4-31b-it": "openrouter",
             "qwen/qwen3-next-80b-a3b-instruct": "openrouter",
             "Qwen/Qwen3-Coder-Next": "openrouter",
             "deepseek/deepseek-v4-flash": "openrouter",

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
+  },
+  "openai/gpt-6-luna": {
+    "input_token_price_per_million": 0.1,
+    "output_token_price_per_million": 0.5
+  },
+  "deepseek/deepseek-v4.1-flash": {
+    "input_token_price_per_million": 0.03,
+    "output_token_price_per_million": 0.6
+  },
+  "deepseek/deepseek-v4-flash-0731": {
+    "input_token_price_per_million": 0.021,
+    "output_token_price_per_million": 0.32
   }
 }

router_inference/config/ainative-router.json

@@ -0,0 +1,15 @@
+{
+  "pipeline_params": {
+    "router_name": "ainative-router",
+    "router_cls_name": "AINativeRouter",
+    "models": [
+      "google/gemma-4-31b-it",
+      "openai/gpt-6-luna",
+      "google/gemini-3-flash-preview",
+      "deepseek/deepseek-v4.1-flash",
+      "qwen/qwen3-235b-a22b-2507",
+      "deepseek/deepseek-v4-flash-0731"
+    ],
+    "description": "AINative Router: content-only routing. The question body (first/last paragraph dropped, field labels and option letters stripped generically) is embedded with all-MiniLM-L6-v2 and classified into 11 content categories by a logistic regression trained only on an external 6,409-query calibration set (same public sources, all RouterArena items removed); each category maps to the model with the best accuracy-cost trade-off on that set (sources weighted equally). No RouterArena template strings, config files, labels or proportions are used. Classifier/policy hash-frozen before routing. Per-row provenance (requested_model, model_used, provider, request_id, invoked_at) recorded."
+  }
+}

router_inference/router/__init__.py

@@ -11,6 +11,7 @@
 from router_inference.router.llm_router import LLMRouter
 from router_inference.router.lynkr_router import LynkrRouter
 from router_inference.router.cruq_sc_router import CruqSCRouter
+from router_inference.router.ainative_router import AINativeRouter
 
 __all__ = [
     "BaseRouter",
@@ -21,4 +22,5 @@
     "ChuzomSoloV32Router",
     "LynkrRouter",
     "CruqSCRouter",
+    "AINativeRouter",
 ]

router_inference/router/ainative_router.py

@@ -0,0 +1,38 @@
+# SPDX-FileCopyrightText: Copyright contributors to the RouterArena project
+# SPDX-License-Identifier: Apache-2.0
+
+"""AINative Router adapter.
+
+Content-only routing: the question body (first/last paragraph dropped, field labels and option letters stripped) is
+classified into a content category by a classifier trained only on external calibration data, and each category maps
+to a model chosen on that data. The router reads no RouterArena files; policy and classifier are hash-frozen
+(artifacts/FREEZE.json) and verified at load.
+"""
+
+import os
+import sys
+
+from router_inference.router.base_router import BaseRouter
+
+_PROJECT_ROOT = os.environ.get(
+    "ARENA_ROUTER_ROOT",
+    os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", "..")),
+)
+if _PROJECT_ROOT not in sys.path:
+    sys.path.insert(0, _PROJECT_ROOT)
+
+
+class AINativeRouter(BaseRouter):
+    def __init__(self, router_name: str):
+        super().__init__(router_name)
+        from arena_router.router import ArenaRouter as _Impl
+
+        self._impl = _Impl()
+        # 정책 선택지는 "model" 또는 "model@<reasoning>" (추론 강도 변형). 예측 파일에는 기본 모델명만 기록하고,
+        # 추론 설정은 각 행의 generated_result.request_params 에 기록된다.
+        unknown = {m.split("@")[0] for m in self._impl.policy.models} - set(self.models)
+        if unknown:
+            raise ValueError(f"policy models not in config: {unknown}")
+
+    def _get_prediction(self, query: str) -> str:
+        return self._impl.route(query).split("@")[0]

universal_model_names.py

@@ -138,6 +138,10 @@
     "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B",
     # KT-ModelRouter pool additions
     "google/gemma-4-31b-it",
+    # AINative Router pool additions (OpenRouter-served)
+    "openai/gpt-6-luna",
+    "deepseek/deepseek-v4.1-flash",
+    "deepseek/deepseek-v4-flash-0731",
 ]