Add Gemini-native judge compatibility to eval harness

eval/config.json

@@ -15,7 +15,7 @@
     "default": {"en": 2850000, "cjk": 850000}
   },
 
-  "_judges_note": "3 non-contestant judges, averaged. The paper used Claude-Sonnet-4.6, Qwen3.5, and Gemini-2.5-Flash. Each entry is an independent OpenAI-compatible endpoint; api_key_env defaults to the top-level one if omitted. List one judge to grade with a single judge (the paper headline is the 3-judge average).",
+  "_judges_note": "3 non-contestant judges, averaged. The paper used Claude-Sonnet-4.6, Qwen3.5, and Gemini-2.5-Flash. Chat endpoints are the default; for gateways exposing Gemini through native contents/candidates payloads, set api_type to gemini_native on that judge. api_key_env defaults to the top-level one if omitted. List one judge to grade with a single judge (the paper headline is the 3-judge average).",
   "judges": [
     {"name": "sonnet",       "base_url": "https://api.openai.com/v1/chat/completions", "model": "claude-sonnet-4-6", "api_key_env": "API_KEY", "max_tokens": 16384},
     {"name": "qwen3.5",      "base_url": "https://api.openai.com/v1/chat/completions", "model": "qwen3.5-plus",       "api_key_env": "API_KEY", "max_tokens": 16384},

eval/run_judge.py

@@ -7,8 +7,10 @@ exclusion). out_of_context_scope answers stay 0 (the model could not ingest the
 
 Panel used in the paper (deliberately models NOT on the leaderboard):
   Claude-Sonnet-4.6 · Qwen3.5 · Gemini-2.5-Flash
-Supply your own judge endpoints in config.json -> "judges" (each OpenAI-compatible). Using a
-single judge is supported (list one) but the paper headline is the 3-judge average.
+Supply your own judge endpoints in config.json -> "judges". OpenAI-compatible chat endpoints are
+the default; gateways that expose Gemini through native `contents`/`candidates` payloads can set
+`"api_type": "gemini_native"` on that judge. Using a single judge is supported (list one) but the
+paper headline is the 3-judge average.
 
 Output: results/<model>.scores.json
   { "per_judge": {judge: {qid: score_0_1}}, "average": {qid: mean_score}, "overall": pct }
@@ -37,6 +39,41 @@ def post(url, key, payload, timeout=180):
     return None
 
 
+def build_payload(judge_cfg, prompt):
+    api_type = judge_cfg.get("api_type", "chat")
+    max_tokens = judge_cfg.get("max_tokens", 16384)
+    temperature = judge_cfg.get("temperature", 0.1)
+    if api_type == "gemini_native":
+        return {
+            "model": judge_cfg["model"],
+            "contents": [{"role": "user", "parts": [{"text": prompt}]}],
+            "generationConfig": {
+                "maxOutputTokens": max_tokens,
+                "temperature": temperature,
+            },
+        }
+    return {
+        "model": judge_cfg["model"],
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": max_tokens,
+        "temperature": temperature,
+    }
+
+
+def response_content(data):
+    content = (data.get("choices", [{}])[0].get("message", {}) or {}).get("content", "")
+    if not content and isinstance(data.get("content"), list):  # Anthropic-native content blocks
+        content = "".join(b.get("text", "") for b in data["content"] if b.get("type") == "text")
+    if not content and isinstance(data.get("candidates"), list):  # Gemini-native candidates
+        parts = []
+        for cand in data["candidates"]:
+            for part in ((cand.get("content") or {}).get("parts") or []):
+                if part.get("text"):
+                    parts.append(part["text"])
+        content = "".join(parts)
+    return content
+
+
 def build_judge_prompt(question, rubric, response):
     if isinstance(rubric, str):
         try: rubric = ast.literal_eval(rubric)
@@ -65,14 +102,10 @@ def parse_total(content):
 def judge_one(judge_cfg, key, question, rubric, response):
     """One judge's score in [0,1], or None on failure. total is a 0-100 sum -> /100, capped at 1."""
     prompt = build_judge_prompt(question, rubric, response)
-    data = post(judge_cfg["base_url"], key,
-                {"model": judge_cfg["model"], "messages": [{"role": "user", "content": prompt}],
-                 "max_tokens": judge_cfg.get("max_tokens", 16384), "temperature": 0.1})
+    data = post(judge_cfg["base_url"], key, build_payload(judge_cfg, prompt))
     if not data:
         return None
-    content = (data.get("choices", [{}])[0].get("message", {}) or {}).get("content", "")
-    if not content and isinstance(data.get("content"), list):  # Anthropic-native content blocks
-        content = "".join(b.get("text", "") for b in data["content"] if b.get("type") == "text")
+    content = response_content(data)
     tot = parse_total(content)
     return None if tot is None else min(tot / 100.0, 1.0)