diff --git a/evaluation/scoring-regressions.json b/evaluation/scoring-regressions.json
index 82663f684..7fe30610d 100644
--- a/evaluation/scoring-regressions.json
+++ b/evaluation/scoring-regressions.json
@@ -21777,6 +21777,264 @@
   "expected_metrics": {
     "stable_unique_record_count": 0
   }
-}
+},
+  {
+    "id": "forbidden-term-longer-label-is-not-error",
+    "category": "format-normalization",
+    "root_cause": "forbidden_phrase_substring_collision",
+    "failure_mode": "false-negative",
+    "case": {
+      "url": "https://example.test/news",
+      "task_kind": "body",
+      "required_facts": [
+        {
+          "alternatives": [
+            "Latest news"
+          ]
+        }
+      ],
+      "forbidden_any": [
+        "Not Found"
+      ],
+      "acceptance_contract": {
+        "version": 1,
+        "schema": "named_link",
+        "minimum_records": 1,
+        "hosts": [
+          "example.test"
+        ],
+        "entry_path": "/story/[^/]+"
+      }
+    },
+    "artifact": "Latest news\n[Story](https://example.test/story/one)\n[Page Not FoundExtra](https://example.test/column)",
+    "ok": true,
+    "expected_task_pass": true,
+    "expected_metrics": {
+      "forbidden_hits": []
+    }
+  },
+  {
+    "id": "forbidden-term-complete-phrase-still-fails",
+    "category": "format-normalization",
+    "root_cause": "forbidden_phrase_substring_collision",
+    "failure_mode": "false-positive",
+    "case": {
+      "url": "https://example.test/news",
+      "task_kind": "body",
+      "required_facts": [
+        {
+          "alternatives": [
+            "Latest news"
+          ]
+        }
+      ],
+      "forbidden_any": [
+        "Not Found"
+      ],
+      "acceptance_contract": {
+        "version": 1,
+        "schema": "named_link",
+        "minimum_records": 1,
+        "hosts": [
+          "example.test"
+        ],
+        "entry_path": "/story/[^/]+"
+      }
+    },
+    "artifact": "Latest news\n[Story](https://example.test/story/one)\nPage Not Found.",
+    "ok": true,
+    "expected_task_pass": false,
+    "expected_metrics": {
+      "forbidden_hits": [
+        "Not Found"
+      ]
+    }
+  },
+  {
+    "id": "forbidden-term-prefixed-word-is-not-error",
+    "category": "format-normalization",
+    "root_cause": "forbidden_phrase_substring_collision",
+    "failure_mode": "false-negative",
+    "case": {
+      "url": "https://example.test/news",
+      "task_kind": "body",
+      "required_facts": [
+        {
+          "alternatives": [
+            "Latest news"
+          ]
+        }
+      ],
+      "forbidden_any": [
+        "Not Found"
+      ],
+      "acceptance_contract": {
+        "version": 1,
+        "schema": "named_link",
+        "minimum_records": 1,
+        "hosts": [
+          "example.test"
+        ],
+        "entry_path": "/story/[^/]+"
+      }
+    },
+    "artifact": "Latest news\n[Story](https://example.test/story/one)\nLatestNot Found",
+    "ok": true,
+    "expected_task_pass": true,
+    "expected_metrics": {
+      "forbidden_hits": []
+    }
+  },
+  {
+    "id": "named-caption-owned-by-link",
+    "category": "structured-extraction",
+    "root_cause": "scoped_markdown_caption_ownership",
+    "failure_mode": "false-negative",
+    "case": {
+      "url": "https://example.test/shows",
+      "task_kind": "body",
+      "acceptance_contract": {
+        "version": 1,
+        "schema": "named_link",
+        "minimum_records": 1,
+        "hosts": [
+          "example.test"
+        ],
+        "entry_path": "/series/[^/]+",
+        "name_validation": "nonempty",
+        "entry_scope": {
+          "html_selector": "main a:has(p)",
+          "markdown_label_pattern": "(?s)<br>\\s*(?P<label>[^<>\\n]+?)\\s*<br>\\s*$"
+        }
+      }
+    },
+    "artifact": "[Artwork<br>Series One<br>](https://example.test/series/one)",
+    "ok": true,
+    "expected_task_pass": true,
+    "expected_metrics": {
+      "stable_record_count": 1
+    }
+  },
+  {
+    "id": "named-caption-nearby-is-not-owned",
+    "category": "structured-extraction",
+    "root_cause": "scoped_markdown_caption_ownership",
+    "failure_mode": "false-positive",
+    "case": {
+      "url": "https://example.test/shows",
+      "task_kind": "body",
+      "acceptance_contract": {
+        "version": 1,
+        "schema": "named_link",
+        "minimum_records": 1,
+        "hosts": [
+          "example.test"
+        ],
+        "entry_path": "/series/[^/]+",
+        "name_validation": "nonempty",
+        "entry_scope": {
+          "html_selector": "main a:has(p)",
+          "markdown_label_pattern": "(?s)<br>\\s*(?P<label>[^<>\\n]+?)\\s*<br>\\s*$"
+        }
+      }
+    },
+    "artifact": "### Series One\n\n[Mehr Infos](https://example.test/series/other)\n\nSeries One<br>\n",
+    "ok": true,
+    "expected_task_pass": false,
+    "expected_metrics": {
+      "stable_record_count": 0
+    }
+  },
+  {
+    "id": "named-caption-unqualified-title-link",
+    "category": "structured-extraction",
+    "root_cause": "scoped_markdown_caption_ownership",
+    "failure_mode": "false-positive",
+    "case": {
+      "url": "https://example.test/shows",
+      "task_kind": "body",
+      "acceptance_contract": {
+        "version": 1,
+        "schema": "named_link",
+        "minimum_records": 1,
+        "hosts": [
+          "example.test"
+        ],
+        "entry_path": "/series/[^/]+",
+        "name_validation": "nonempty",
+        "entry_scope": {
+          "html_selector": "main a:has(p)",
+          "markdown_label_pattern": "(?s)<br>\\s*(?P<label>[^<>\\n]+?)\\s*<br>\\s*$"
+        }
+      }
+    },
+    "artifact": "### [Series One](https://example.test/series/one)",
+    "ok": true,
+    "expected_task_pass": false,
+    "expected_metrics": {
+      "stable_record_count": 0
+    }
+  },
+  {
+    "id": "forbidden-cjk-remains-unsegmented",
+    "category": "format-normalization",
+    "root_cause": "forbidden_phrase_substring_collision",
+    "failure_mode": "false-positive",
+    "case": {
+      "url": "https://example.test/news",
+      "task_kind": "body",
+      "forbidden_any": [
+        "访问受限"
+      ]
+    },
+    "artifact": "当前页面访问受限，请稍后重试",
+    "ok": true,
+    "expected_task_pass": false,
+    "expected_metrics": {
+      "forbidden_hits": [
+        "访问受限"
+      ]
+    }
+  },
+  {
+    "id": "forbidden-url-path-is-not-visible",
+    "category": "format-normalization",
+    "root_cause": "forbidden_phrase_substring_collision",
+    "failure_mode": "false-negative",
+    "case": {
+      "url": "https://example.test/news",
+      "task_kind": "body",
+      "forbidden_any": [
+        "Not Found"
+      ]
+    },
+    "artifact": "[Latest news](https://example.test/not-found)",
+    "ok": true,
+    "expected_task_pass": true,
+    "expected_metrics": {
+      "forbidden_hits": []
+    }
+  },
+  {
+    "id": "forbidden-phrase-case-and-markup",
+    "category": "format-normalization",
+    "root_cause": "forbidden_phrase_substring_collision",
+    "failure_mode": "false-positive",
+    "case": {
+      "url": "https://example.test/news",
+      "task_kind": "body",
+      "forbidden_any": [
+        "Not Found"
+      ]
+    },
+    "artifact": "**NOT** FOUND: requested page is unavailable.",
+    "ok": true,
+    "expected_task_pass": false,
+    "expected_metrics": {
+      "forbidden_hits": [
+        "Not Found"
+      ]
+    }
+  }
 ]
 }
diff --git a/harness/web_fetch/metrics.py b/harness/web_fetch/metrics.py
index 2b5ddaa30..62680495f 100644
--- a/harness/web_fetch/metrics.py
+++ b/harness/web_fetch/metrics.py
@@ -441,6 +441,15 @@ def fact_in_text(text: str, fact: str) -> bool:
     return False
 
 
+def forbidden_in_text(text: str, phrase: str) -> bool:
+    """Match ASCII word boundaries without segmenting CJK phrases."""
+    if not phrase:
+        return False
+    left = r'(?<![a-z0-9_])' if phrase[0].isascii() and phrase[0].isalnum() else ''
+    right = r'(?![a-z0-9_])' if phrase[-1].isascii() and phrase[-1].isalnum() else ''
+    return bool(re.search(left + re.escape(phrase) + right, text))
+
+
 def text_corruption(value: str) -> tuple[int, float, bool]:
     """Detect substantial UTF-8 replacement-character damage in text output."""
     count = value.count('\ufffd')
@@ -534,12 +543,12 @@ def evaluate_content(
         ]
         block_hits = [hit for hit in block_hits if hit not in block_reference_exemptions]
     reference_exemptions = [item for item in (forbidden_any or [])
-                            if reference_folded and (text := normalize_for_match(item)) and text in reference_folded]
+                            if reference_folded and forbidden_in_text(reference_folded, normalize_for_match(item))]
     if acceptance_contract:
         # A historical snapshot cannot exempt an explicit stable-mode block rule.
         reference_exemptions = []
     forbidden = [item for item in (forbidden_any or [])
-                 if item not in reference_exemptions and (text := normalize_for_match(item)) and text in folded]
+                 if item not in reference_exemptions and forbidden_in_text(folded, normalize_for_match(item))]
     lines = [re.sub(r"\s+", " ", line).strip().casefold() for line in answer_text.splitlines()]
     lines = [line for line in lines if line]
     unique_line_ratio = round(len(set(lines)) / len(lines), 4) if lines else 1.0
diff --git a/tests/test_scoring_regressions.py b/tests/test_scoring_regressions.py
index e816ae3d7..e56f7aaa2 100644
--- a/tests/test_scoring_regressions.py
+++ b/tests/test_scoring_regressions.py
@@ -95,6 +95,8 @@ def test_every_registered_false_verdict_control_runs_through_shared_scorer(self)
         paired.add("declared_full_page_stable_text_scope")
         paired.add("inline_emphasis_in_structured_fields")
         paired.add("markdown_list_prefix_ownership")
+        paired.add("forbidden_phrase_substring_collision")
+        paired.add("scoped_markdown_caption_ownership")
         for root_cause in paired:
             self.assertEqual(root_modes.get(root_cause), {"false-positive", "false-negative"})