Add Compass CoreRouter v2 submission (compass-corerouter; supersedes #237)

llm_inference/model_inference.py

@@ -12,6 +12,7 @@
 import logging
 from typing import Dict, Any
 from openai import OpenAI
+import re
 import tiktoken
 
 logger = logging.getLogger(__name__)
@@ -194,6 +195,7 @@ def _get_provider(self, model_name: str) -> str:
             "qwen/qwen3-next-80b-a3b-instruct": "openrouter",
             "Qwen/Qwen3-Coder-Next": "openrouter",
             "deepseek/deepseek-v4-flash": "openrouter",
+            "deepseek/deepseek-v4.1-flash-1017-reasoning-low": "openrouter",
             "x-ai/grok-4.1-fast": "openrouter",
             "mistralai/devstral-2512:free": "openrouter",
             "meta-llama/llama-3.3-70b-instruct": "openrouter",
@@ -321,8 +323,20 @@ def _call_openrouter(self, model_name: str, prompt: str) -> Dict[str, Any]:
             api_key=openrouter_api_key,
         )
 
+        # "<model>-reasoning-<effort>" selects the same model with an explicit
+        # OpenRouter reasoning effort (e.g. deepseek/deepseek-v4.1-flash-1017-reasoning-low).
+        api_model, extra_body = model_name, None
+        match = re.fullmatch(
+            r"(.+)-reasoning-(none|minimal|low|medium|high)", model_name
+        )
+        if match:
+            api_model = match.group(1)
+            extra_body = {"reasoning": {"effort": match.group(2)}}
+
         response = client.chat.completions.create(
-            model=model_name, messages=[{"role": "user", "content": prompt}]
+            model=api_model,
+            messages=[{"role": "user", "content": prompt}],
+            extra_body=extra_body,
         )
 
         usage = getattr(response, "usage", None)

model_cost/model_cost.json

@@ -386,5 +386,9 @@
   "google/gemma-4-31b-it": {
     "input_token_price_per_million": 0.08,
     "output_token_price_per_million": 0.35
+  },
+  "deepseek/deepseek-v4.1-flash-1017-reasoning-low": {
+    "input_token_price_per_million": 0.15,
+    "output_token_price_per_million": 0.6
   }
 }

router_inference/config/compass-corerouter.json

@@ -0,0 +1,18 @@
+{
+  "pipeline_params": {
+    "router_name": "compass-corerouter",
+    "router_cls_name": "CompassCoreRouter",
+    "models": [
+      "qwen/qwen3-235b-a22b-2507",
+      "google/gemma-4-31b-it",
+      "gemini-3-flash-preview",
+      "deepseek/deepseek-v4-flash",
+      "qwen/qwen3-next-80b-a3b-instruct",
+      "deepseek/deepseek-v4.1-flash-1017-reasoning-low"
+    ],
+    "router_version": "corerouter-v2-hybrid-agreement-critic-veto-reasoning-tier-prompt-only",
+    "routing_input": "prompt text only: a TF-IDF + logistic-regression task classifier trained on our own practice prompts predicts the task type",
+    "policy": "per-task model pick fitted on our own practice data (5-fold CV, lambda 0.02); where the pick is gemini-3-flash-preview, Qwen3-235B and Gemma-4-31B answer first; Gemini is called when their final answers disagree, or when they agree but a self-hosted answer critic (fine-tuned on our own practice data; it reads only the prompt and our models' answers) scores the shared answer below 0.5; for the MMLU-Pro task type the escalation model is DeepSeek V4.1 Flash 1017 with reasoning effort low instead of Gemini, and the AIME and MATH task types go to it directly (choices fitted by the same cross-validation on our own practice data)",
+    "cost_accounting": "every call is charged on the returned row: other calls' dollar cost is added as returned-model input-price tokens; see generated_result.cascade_accounting"
+  }
+}