From a4bbbd4d03b8fd89a29456107f6b970020c05860 Mon Sep 17 00:00:00 2001
From: keeponlight <acm@hgui.org>
Date: Fri, 23 Oct 2026 23:28:36 +0800
Subject: [PATCH] fix(prompt): use a neutral label for top-program metrics

---
 openevolve/prompts/defaults/fragments.json |  2 +-
 tests/test_prompt_sampler.py               | 54 ++++++++++++++++++++++
 2 files changed, 55 insertions(+), 1 deletion(-)

diff --git a/openevolve/prompts/defaults/fragments.json b/openevolve/prompts/defaults/fragments.json
index b59410405a..0ecfba065b 100644
--- a/openevolve/prompts/defaults/fragments.json
+++ b/openevolve/prompts/defaults/fragments.json
@@ -22,7 +22,7 @@
   "attempt_all_metrics_improved": "Improvement in all metrics",
   "attempt_all_metrics_regressed": "Regression in all metrics",
   "attempt_mixed_metrics": "Mixed results",
-  "top_program_metrics_prefix": "Performs well on",
+  "top_program_metrics_prefix": "Metric:",
   "diverse_program_metrics_prefix": "Alternative approach to",
   "inspiration_type_diverse": "Diverse",
   "inspiration_type_migrant": "Migrant",
diff --git a/tests/test_prompt_sampler.py b/tests/test_prompt_sampler.py
index fc29e97306..03374617e8 100644
--- a/tests/test_prompt_sampler.py
+++ b/tests/test_prompt_sampler.py
@@ -2,7 +2,11 @@
 Tests for PromptSampler in openevolve.prompt.sampler
 """
 
+import json
+import tempfile
 import unittest
+from pathlib import Path
+
 from openevolve.config import Config
 from openevolve.prompt.sampler import PromptSampler
 
@@ -49,6 +53,56 @@ def test_build_prompt(self):
         # Check that the score value appears in the prompt (either as fitness or metrics)
         self.assertIn("0.5", prompt["user"])
 
+    def test_top_program_metrics_use_neutral_descriptions(self):
+        """Report metric values without inferring that they indicate good performance."""
+        cases = [
+            ("score", 0.0, "0.0000"),
+            ("score", -1.0, "-1.0000"),
+            ("error", "evaluation failed", "evaluation failed"),
+            ("latency", 10000.0, "10000.0000"),
+        ]
+        for name, value, formatted_value in cases:
+            with self.subTest(metric=name, value=value):
+                prompt = self.prompt_sampler.build_prompt(
+                    top_programs=[{"id": "top1", "metrics": {name: value}}],
+                )
+
+                self.assertIn(f"Key features: Metric: {name} ({formatted_value})", prompt["user"])
+                self.assertNotIn("Performs well on", prompt["user"])
+
+    def test_top_program_explicit_key_features_are_preserved(self):
+        """Keep caller-provided descriptions instead of generating metric descriptions."""
+        prompt = self.prompt_sampler.build_prompt(
+            top_programs=[
+                {
+                    "id": "top1",
+                    "metrics": {"score": 0.0},
+                    "key_features": ["Uses memoization", "Avoids repeated work"],
+                }
+            ],
+        )
+
+        self.assertIn("Key features: Uses memoization, Avoids repeated work", prompt["user"])
+        self.assertNotIn("Metric: score", prompt["user"])
+
+    def test_top_program_metrics_prefix_can_be_overridden(self):
+        """Load a custom metric prefix through the configured template directory."""
+        with tempfile.TemporaryDirectory() as template_dir:
+            fragments_path = Path(template_dir) / "fragments.json"
+            fragments_path.write_text(
+                json.dumps({"top_program_metrics_prefix": "Observed:"}), encoding="utf-8"
+            )
+            config = Config()
+            config.prompt.template_dir = template_dir
+            sampler = PromptSampler(config.prompt)
+
+            prompt = sampler.build_prompt(
+                top_programs=[{"id": "top1", "metrics": {"score": 0.0}}],
+            )
+
+        self.assertIn("Key features: Observed: score (0.0000)", prompt["user"])
+        self.assertNotIn("Metric: score", prompt["user"])
+
 
 if __name__ == "__main__":
     unittest.main()