diff --git a/hermes-agent/leoclean-bin/kb_tool.py b/hermes-agent/leoclean-bin/kb_tool.py
index 67b01e4..f8df56f 100755
--- a/hermes-agent/leoclean-bin/kb_tool.py
+++ b/hermes-agent/leoclean-bin/kb_tool.py
@@ -468,6 +468,22 @@ def _has_unnegated_action(query: str, actions: tuple[str, ...]) -> bool:
     return False
 
 
+def _has_unnegated_source_action(query: str) -> bool:
+    """Detect source-intake verbs without treating a prohibition as intent."""
+
+    action_pattern = re.compile(
+        r"\b(?:send|hand|drop|upload|learn|absorb|ingest|extract|stage|add)(?:s|ed|ing)?\b|"
+        r"\bmake it part\b"
+    )
+    for segment in re.split(r"(?<=[.!?;])\s+|\n+", query.lower()):
+        if not action_pattern.search(segment):
+            continue
+        if "read-only" in segment or re.search(r"\b(?:do not|don't|dont|must not|without)\b", segment):
+            continue
+        return True
+    return False
+
+
 def _has_unnegated_database_state_request(query: str) -> bool:
     """Detect an actual DB-state question instead of incidental lifecycle vocabulary."""
 
@@ -705,9 +721,16 @@ def operational_contracts(
     candidate_claim_composition_question = bool(
         any(
             term in lowered
-            for term in ("candidate claim", "reviewable candidate", "candidate only", "revised exact claim body")
+            for term in (
+                "candidate claim",
+                "reviewable candidate",
+                "candidate only",
+                "revised exact claim body",
+                "review-only proposal",
+                "review only proposal",
+            )
         )
-        and any(term in lowered for term in ("draft", "propose", "revise", "rewrite", "narrow", "return"))
+        and any(term in lowered for term in ("draft", "propose", "revise", "rewrite", "narrow", "return", "turn"))
         and not existing_claim_audit_question
     )
     claim_evidence_challenge_question = bool(
@@ -826,13 +849,7 @@ def operational_contracts(
     )
     source_terms = ("document", "pdf", "tweet", "attachment", "ingest", "intake", "absorb", "extract")
     broad_source_object = bool(re.search(r"\b(?:report|article|file|url|link)s?\b", normalized))
-    broad_source_action = bool(
-        re.search(
-            r"\b(?:send|hand|drop|upload|learn|absorb|ingest|extract|stage|add)(?:s|ed|ing)?\b|"
-            r"\bmake it part\b",
-            normalized,
-        )
-    )
+    broad_source_action = _has_unnegated_source_action(query)
     source_intake_question = any(term in lowered for term in source_terms) or (
         broad_source_object and broad_source_action
     )
diff --git a/tests/test_hermes_leoclean_kb_bridge_source.py b/tests/test_hermes_leoclean_kb_bridge_source.py
index bafa184..89aff9c 100644
--- a/tests/test_hermes_leoclean_kb_bridge_source.py
+++ b/tests/test_hermes_leoclean_kb_bridge_source.py
@@ -1968,6 +1968,15 @@ def test_vps_bridge_direct_intents_reject_unrelated_and_resolve_overlaps() -> No
     }
     assert reasoning_followup_ids & direct_ids == set()
 
+    review_only_proposal_ids = {
+        item["id"]
+        for item in module.operational_contracts(
+            "Turn the surviving claim into a review-only proposal: exact body, confidence, falsifier, required "
+            "sources, and links to claims that support or contradict it. Do not stage or write anything."
+        )
+    }
+    assert review_only_proposal_ids == {"reply_budget"}
+
     chat_memory_ids = {
         item["id"]
         for item in module.operational_contracts(