diff --git a/agentrace/checks.py b/agentrace/checks.py
index 4204050..5f65b73 100644
--- a/agentrace/checks.py
+++ b/agentrace/checks.py
@@ -115,6 +115,50 @@ def check_destructive_command(run: AgentRun) -> list[Finding]:
     return []
 
 
+def check_risky_verdict(run: AgentRun) -> list[Finding]:
+    """A confident operational recommendation with no supporting artifact."""
+    action = r"(?:delet(?:e|ed|ing)|drop(?:ped|ping)?|deploy(?:ed|ing)?|merg(?:e|ed|ing)|revok(?:e|ed|ing)|force[- ]push(?:ed|ing)?)"
+    verdict_patterns = [
+        rf"\b(?:safe|okay|ok|fine|reasonable|appropriate|approved?|ready)\s+to\s+(?:\w+\s+){{0,3}}{action}\b",
+        rf"\b(?:go ahead(?: and)?|proceed(?: with)?|recommend(?:ed|ation)?(?: is|:)?|you (?:can|should|may))\s+(?:\w+\s+){{0,3}}{action}\b",
+        rf"\b{action}\b(?:\s+\w+){{0,4}}\s+(?:is|looks?)\s+(?:safe|okay|fine|approved?|ready)\b",
+        rf"\b{action}\b[^.\n]{{0,80}}\bgo ahead\b",
+    ]
+    evidence_patterns = [
+        r"```[\s\S]*?```",
+        r"`[^`\n]+`",
+        r"(?m)^\s{4,}\S",
+        r"(?m)^(?:diff --git |Index: |@@\s|[+-]{3}\s+\S)",
+        r"(?<![\w/])(?:\.{1,2}/|/[^\s]+|[\w.-]+/[\w./-]+)(?::\d+)?\b",
+        r"(?<![\w/])(?:[\w-]+\.)+[A-Za-z][A-Za-z0-9_-]{0,7}\b",
+        r"(?<![\w/])(?:[\w.-]+/)*[\w-]+\.[A-Za-z][A-Za-z0-9_-]{0,7}:\d+\b",
+        r"[\"“'](?=[^\"\n“”']{0,240}\b(?:SELECT|INSERT|UPDATE|DELETE\s+FROM|DROP\s+TABLE|pytest|npm\s+test|make\s+test|passed|failed|rows?|exit\s+code|HTTP\s+\d{3})\b)[^\"\n“”']+[\"”']",
+    ]
+    verdict = None
+    for pattern in verdict_patterns:
+        match = re.search(pattern, run.result, re.IGNORECASE)
+        if match:
+            verdict = match
+            break
+    if not verdict:
+        return []
+
+    preceding = run.result[max(0, verdict.start() - 40) : verdict.start()]
+    if re.search(r"\b(?:do not|don't|cannot|can't|should not|shouldn't|never|avoid|not)\b", preceding, re.IGNORECASE):
+        return []
+    if any(re.search(pattern, run.result, re.IGNORECASE) for pattern in evidence_patterns):
+        return []
+
+    return [
+        Finding(
+            "risky_verdict",
+            "medium",
+            "Risky operational verdict has no supporting file, code, command, or test evidence.",
+            _context(run.result, verdict.start()),
+        )
+    ]
+
+
 def check_unverified_claim(run: AgentRun) -> list[Finding]:
     """Hedged language presented as a finding.
 
@@ -326,6 +370,7 @@ def check_runaway(run: AgentRun, slow_s: float = 900.0) -> list[Finding]:
     check_empty_result,
     check_refused_or_gave_up,
     check_destructive_command,
+    check_risky_verdict,
     check_unverified_claim,
     check_absence_as_evidence,
     check_self_contradiction,
diff --git a/tests/test_risky_verdict.py b/tests/test_risky_verdict.py
new file mode 100644
index 0000000..3a3cbe0
--- /dev/null
+++ b/tests/test_risky_verdict.py
@@ -0,0 +1,75 @@
+from datetime import UTC, datetime
+
+import pytest
+
+from agentrace.checks import analyse
+from agentrace.parse import AgentRun
+
+
+def _run(result: str) -> AgentRun:
+    return AgentRun(
+        tool_use_id="toolu_risky_verdict_test",
+        description="test run",
+        prompt="Review the operation and report what should be done.",
+        result=result,
+        started_at=datetime(2026, 7, 16, 12, 0, tzinfo=UTC),
+        ended_at=datetime(2026, 7, 16, 12, 1, tzinfo=UTC),
+    )
+
+
+def _findings(result: str):
+    return [f for f in analyse(_run(result)) if f.check == "risky_verdict"]
+
+
+@pytest.mark.parametrize(
+    "verdict",
+    [
+        "It is safe to delete the table, go ahead",
+        "It is safe to drop the table, go ahead",
+        "It is safe to deploy the change, go ahead",
+        "It is safe to merge the branch, go ahead",
+        "It is safe to revoke the token, go ahead",
+        "It is safe to force-push the branch, go ahead",
+    ],
+)
+def test_risky_operational_verdict_without_evidence_is_flagged(verdict):
+    findings = _findings(verdict)
+    assert findings
+    assert findings[0].severity == "medium"
+
+
+def test_risky_verdict_with_file_path_is_clean():
+    assert not _findings("Reviewed db/migrations/2026_drop_users.sql; it is safe to drop the table.")
+
+
+def test_risky_verdict_with_file_line_evidence_is_clean():
+    assert not _findings("src/db/schema.sql:42 confirms it is safe to drop the table.")
+
+
+def test_risky_verdict_with_quoted_query_output_is_clean():
+    assert not _findings('"SELECT count(*) FROM users returned 0 rows"; it is safe to delete the table.')
+
+
+def test_risky_verdict_with_code_span_is_clean():
+    assert not _findings("`pytest tests/test_migrations.py -q` passed, so it is safe to deploy.")
+
+
+def test_risky_verdict_with_diff_evidence_is_clean():
+    result = "@@ -10,2 +10,2 @@\n- old setting\n+ new setting\nIt is safe to deploy."
+    assert not _findings(result)
+
+
+def test_version_number_is_not_treated_as_file_evidence():
+    assert _findings("It is safe to deploy v1.2.3, go ahead")
+
+
+@pytest.mark.parametrize(
+    "result",
+    [
+        "The merge function combines two sorted lists.",
+        "Delete is a SQL keyword documented in the parser reference.",
+        "The deployment section explains how merge queues work.",
+    ],
+)
+def test_ordinary_action_words_are_clean(result):
+    assert not _findings(result)