diff --git a/wizard-work/spark/verify_csv.py b/wizard-work/spark/verify_csv.py
index 277c099..0d69954 100644
--- a/wizard-work/spark/verify_csv.py
+++ b/wizard-work/spark/verify_csv.py
@@ -1,24 +1,40 @@
 #!/usr/bin/env python3
-"""Verify CSV output is valid."""
+"""Verify CSV output is valid.
+
+Exit-safe since 2026-10-27 (Q63's STRIPPED_ASSERT class, found by
+volt's Q65 census pass at `c64e39d`): every check below was an `assert`, so
+`python3 -O` -- which compiles asserts away -- printed PASS and exited 0
+against a wrong CSV. `assert` is not a failure path; `_check` raises.
+The subject and the messages are unchanged; the only difference is that a
+failing check now survives -O.
+"""
 
 import csv
 import json
+import sys
+
+
+def _check(cond, message):
+    if not cond:
+        print("FAIL: " + message, file=sys.stderr)
+        raise SystemExit(1)
+
 
 # Verify CSV is valid
 with open('/root/.hermes/wizard-outputs/spark/output.csv', 'r') as f:
     reader = csv.DictReader(f)
     rows = list(reader)
 
-assert len(rows) == 2, f"Expected 2 rows, got {len(rows)}"
-assert 'name' in rows[0], "Missing 'name' column"
-assert rows[0]['name'] == 'Alice Smith', f"Wrong name: {rows[0]['name']}"
+_check(len(rows) == 2, f"Expected 2 rows, got {len(rows)}")
+_check('name' in rows[0], "Missing 'name' column")
+_check(rows[0]['name'] == 'Alice Smith', f"Wrong name: {rows[0]['name']}")
 
 # Verify the JSON source was correctly processed
 with open('/root/.hermes/wizard-outputs/spark/sample_data.json', 'r') as f:
     source = json.load(f)
 
-assert len(source) == 2, "Source JSON should have 2 records"
+_check(len(source) == 2, "Source JSON should have 2 records")
 
 print("PASS: CSV is valid with correct data structure")
 print(f"Records: {len(rows)}")
-print(f"Columns: {sorted(rows[0].keys())}")
\ No newline at end of file
+print(f"Columns: {sorted(rows[0].keys())}")