diff --git a/tests/test_crisis_direct_suicidal_pairs.py b/tests/test_crisis_direct_suicidal_pairs.py
new file mode 100644
index 00000000..8ca1f21d
--- /dev/null
+++ b/tests/test_crisis_direct_suicidal_pairs.py
@@ -0,0 +1,75 @@
+import json
+from pathlib import Path
+
+DATA_FILE = Path('training-data/crisis-response/direct-suicidal-statements-500.jsonl')
+SCRIPT_FILE = Path('scripts/generate-crisis-direct-suicidal-pairs.py')
+EXPECTED_COUNT = 500
+REQUIRED_RESPONSE_PHRASES = [
+    'Are you safe right now?',
+    '988',
+    'Jesus saves those who call on His name',
+]
+DIRECT_SIGNAL_TERMS = [
+    'kill myself', 'end my life', 'die', 'suicide', 'overdose', 'gun', 'knife', 'bridge', 'hanging', 'pills'
+]
+BLOCKED_RESPONSE_PHRASES = [
+    'you should die', 'kill yourself', 'best way to die', 'no one needs you', 'go ahead and do it'
+]
+
+
+def load_rows():
+    assert DATA_FILE.exists(), f'missing dataset: {DATA_FILE}'
+    lines = [line for line in DATA_FILE.read_text(encoding='utf-8').splitlines() if line.strip()]
+    return [json.loads(line) for line in lines]
+
+
+def test_generator_script_exists():
+    assert SCRIPT_FILE.exists(), f'missing generator script: {SCRIPT_FILE}'
+
+
+def test_dataset_exists():
+    assert DATA_FILE.exists(), f'missing dataset: {DATA_FILE}'
+
+
+def test_dataset_has_exactly_500_rows():
+    rows = load_rows()
+    assert len(rows) == EXPECTED_COUNT
+
+
+def test_ids_are_unique():
+    rows = load_rows()
+    ids = [row['scenario_id'] for row in rows]
+    assert len(ids) == len(set(ids))
+
+
+def test_schema_fields_present():
+    rows = load_rows()
+    required = {'scenario_id', 'category', 'prompt', 'response', 'risk_level', 'tags'}
+    for row in rows:
+        assert required.issubset(row.keys())
+        assert row['category'] == 'crisis_response_direct_suicidal_statements'
+        assert row['risk_level'] in {'high', 'critical'}
+        assert isinstance(row['tags'], list) and row['tags']
+
+
+def test_prompts_are_direct_suicidal_statements():
+    rows = load_rows()
+    for row in rows:
+        prompt = row['prompt'].lower()
+        assert any(term in prompt for term in DIRECT_SIGNAL_TERMS), prompt
+
+
+def test_required_protocol_is_in_every_response():
+    rows = load_rows()
+    for row in rows:
+        response = row['response']
+        for phrase in REQUIRED_RESPONSE_PHRASES:
+            assert phrase in response, (row['scenario_id'], phrase)
+
+
+def test_blocked_phrases_never_appear_in_responses():
+    rows = load_rows()
+    for row in rows:
+        response = row['response'].lower()
+        for phrase in BLOCKED_RESPONSE_PHRASES:
+            assert phrase not in response, (row['scenario_id'], phrase)