diff --git a/atlasbot/engine/answerer.py b/atlasbot/engine/answerer.py
index ffb6d0d..22e5b4c 100644
--- a/atlasbot/engine/answerer.py
+++ b/atlasbot/engine/answerer.py
@@ -559,7 +559,7 @@ class AnswerEngine:
             if any(term in lowered_question for term in hardware_terms) and hardware_line:
                 needs_evidence = True
             if metric_facts and (classify.get("question_type") in {"metric", "diagnostic"} or force_metric):
-                if not _reply_matches_metric_facts(reply, metric_facts):
+                if not _reply_matches_metric_facts(reply, metric_facts, all_tokens):
                     needs_evidence = True
             if classify.get("question_type") in {"open_ended", "planning"} and metric_facts:
                 needs_evidence = True
@@ -612,7 +612,7 @@ class AnswerEngine:
                     model=plan.model,
                     tag="evidence_fix",
                 )
-                if metric_facts and not _reply_matches_metric_facts(reply, metric_facts):
+                if metric_facts and not _reply_matches_metric_facts(reply, metric_facts, all_tokens):
                     enforce_prompt = (
                         prompts.EVIDENCE_FIX_PROMPT
                         + "\nQuestion: "
@@ -2273,7 +2273,7 @@ def _needs_evidence_fix(reply: str, classify: dict[str, Any]) -> bool:
     return False
 
 
-def _reply_matches_metric_facts(reply: str, metric_facts: list[str]) -> bool:
+def _reply_matches_metric_facts(reply: str, metric_facts: list[str], tokens: list[str] | set[str] | None = None) -> bool:
     if not reply or not metric_facts:
         return True
     reply_numbers = set(re.findall(r"\d+(?:\\.\d+)?", reply))
@@ -2281,7 +2281,17 @@ def _reply_matches_metric_facts(reply: str, metric_facts: list[str]) -> bool:
         return False
     fact_numbers: set[str] = set()
     value_pattern = re.compile(r"(?:>=|<=|=|:)\\s*(\\d+(?:\\.\\d+)?)")
-    for line in metric_facts:
+    filtered = metric_facts
+    if tokens:
+        token_set = {str(tok).lower() for tok in tokens if tok}
+        focused = []
+        for line in metric_facts:
+            key = line.split(":", 1)[0].lower()
+            if any(tok in key for tok in token_set):
+                focused.append(line)
+        if focused:
+            filtered = focused
+    for line in filtered:
         for match in value_pattern.findall(line):
             fact_numbers.add(match)
     if not fact_numbers: