diff --git a/atlasbot/engine/answerer.py b/atlasbot/engine/answerer.py
index 91ed684..e557f7a 100644
--- a/atlasbot/engine/answerer.py
+++ b/atlasbot/engine/answerer.py
@@ -1005,6 +1005,8 @@ def _metric_candidate_lines(lines: list[str], keywords: list[str] | None, limit:
     candidates: list[str] = []
     for line in lines:
         line_lower = line.lower()
+        if line_lower.startswith("lexicon_") or line_lower.startswith("units:"):
+            continue
         if prefer_node and "pod_" in line_lower:
             continue
         if "hottest:" in line_lower:
diff --git a/atlasbot/llm/prompts.py b/atlasbot/llm/prompts.py
index f5a2fb8..590c915 100644
--- a/atlasbot/llm/prompts.py
+++ b/atlasbot/llm/prompts.py
@@ -218,6 +218,7 @@ FACT_SELECT_PROMPT = (
     "If the question asks for highest/hottest and Candidates include a line starting with 'hottest:', you must include that line. "
     "If the question mentions nodes and a 'hottest:' line exists, prefer node-level facts over pod-level lines. "
     "Avoid pod_* lines unless the question explicitly mentions pods. "
+    "Exclude lexicon/definition lines; choose lines with concrete numeric values. "
     "Return JSON with field: lines (list of strings). If none apply, return {{\"lines\": []}}."
 )