fix: resolve LLM judge overall_score overriding weighted benchmark score

src/imagent_bench/scoring.py

@@ -1,4 +1,4 @@
-from __future__ import annotations
+from __future__ import annotations
 
 import base64
 import html
@@ -125,8 +125,6 @@ def evaluate_openrouter_vision(image_path: Path, *, prompt: str, config: dict[st
     parsed = _parse_judge_json(content)
     dimension_scores = _normalize_dimension_scores(parsed.get("scores", {}), dimensions)
     overall_score = _weighted_score(dimension_scores, dimensions)
-    if isinstance(parsed.get("overall_score"), int | float):
-        overall_score = _clamp_score(float(parsed["overall_score"]))
 
     usage = response_payload.get("usage")
     usage = usage if isinstance(usage, dict) else {}

tests/test_scoring.py

@@ -59,3 +59,56 @@ def read(self) -> bytes:
     assert result["dimensions"] == {"prompt_alignment": 90.0, "visual_quality": 80.0}
     assert result["cost_usd"] == 0.002
     assert result["judge"]["model"] == "judge/model"
+
+
+def test_openrouter_vision_judge_ignores_self_reported_overall_score(
+    monkeypatch, tmp_path: Path
+) -> None:  # noqa: ANN001
+    image_path = tmp_path / "image.png"
+    image_path.write_bytes(b"fake-image")
+
+    class FakeResponse:
+        def __enter__(self):  # noqa: ANN001
+            return self
+
+        def __exit__(self, *args):  # noqa: ANN002
+            return None
+
+        def read(self) -> bytes:
+            return json.dumps(
+                {
+                    "model": "judge/model",
+                    "choices": [
+                        {
+                            "message": {
+                                "content": json.dumps(
+                                    {
+                                        "scores": {
+                                            "prompt_alignment": 40,
+                                            "visual_quality": 40,
+                                        },
+                                        "overall_score": 99,
+                                        "rationale": "inflated self-report",
+                                    }
+                                )
+                            }
+                        }
+                    ],
+                }
+            ).encode("utf-8")
+
+    monkeypatch.setenv("OPENROUTER_API_KEY", "test-key")
+    monkeypatch.setattr("urllib.request.urlopen", lambda *args, **kwargs: FakeResponse())
+
+    result = evaluate_openrouter_vision(
+        image_path,
+        prompt="Create a clean product image.",
+        config={
+            "provider": "openrouter_vision",
+            "model": "judge/model",
+            "dimensions": {"prompt_alignment": 0.75, "visual_quality": 0.25},
+        },
+    )
+
+    assert result["overall_score"] == 40.0
+    assert result["judge"]["raw_response"]["overall_score"] == 99