diff --git a/backend/tests/test_yolo_training_supervisor.py b/backend/tests/test_yolo_training_supervisor.py
index 0b11596d..fc35cd3a 100644
--- a/backend/tests/test_yolo_training_supervisor.py
+++ b/backend/tests/test_yolo_training_supervisor.py
@@ -33,5 +33,17 @@ def test_container_running_fails_closed_on_inspect_error(monkeypatch) -> None:
 def test_completion_command_requires_non_empty_string_list(tmp_path: Path) -> None:
     invalid = tmp_path / "command.json"
     invalid.write_text('{"shell": "unsafe"}', encoding="utf-8")
-    value = MODULE.json.loads(invalid.read_text(encoding="utf-8"))
-    assert not isinstance(value, list)
+    try:
+        MODULE.load_completion_command(invalid)
+    except ValueError as exc:
+        assert "non-empty JSON list" in str(exc)
+    else:
+        raise AssertionError("mapping-shaped completion command was accepted")
+
+
+def test_completion_command_preserves_exact_argv_without_shell_parsing(tmp_path: Path) -> None:
+    command = tmp_path / "command.json"
+    command.write_text('["docker", "exec", "-d", "geointel", "python", "loop.py"]', encoding="utf-8")
+    assert MODULE.load_completion_command(command) == [
+        "docker", "exec", "-d", "geointel", "python", "loop.py"
+    ]
diff --git a/docs/CODEX_EXECUTION_LOG.md b/docs/CODEX_EXECUTION_LOG.md
index 8288b56d..10f291dc 100644
--- a/docs/CODEX_EXECUTION_LOG.md
+++ b/docs/CODEX_EXECUTION_LOG.md
@@ -11799,6 +11799,11 @@ Deployment evidence:
   `results.png` appears. The bound command permits up to 20 iterations, keeps
   protected test/background closed until calibration passes, and preserves
   the exact AdamW, 180-degree rotation, flip, scale and translation contract.
+- Extracted and tested strict completion-command validation. Only a non-empty
+  JSON list of non-empty argv strings is accepted; mappings, empty arguments,
+  malformed JSON and filesystem errors produce an explicit fail-closed
+  supervisor state before any subprocess is started. The updated supervisor
+  was activated live with one remaining v37 GPU process.
 - Added those aerial augmentation parameters to the orchestrator CLI and
   training command, preventing later failure-driven checkpoints from silently
   reverting to generic orientation assumptions.
diff --git a/scripts/supervise_container_yolo_training.py b/scripts/supervise_container_yolo_training.py
index 4e7716e1..c4e98fe1 100644
--- a/scripts/supervise_container_yolo_training.py
+++ b/scripts/supervise_container_yolo_training.py
@@ -37,6 +37,13 @@ def write_state(path: Path, payload: dict) -> None:
     temporary.replace(path)
 
 
+def load_completion_command(path: Path) -> list[str]:
+    command = json.loads(path.read_text(encoding="utf-8"))
+    if not isinstance(command, list) or not command or not all(isinstance(x, str) and x for x in command):
+        raise ValueError("completion command must be a non-empty JSON list of non-empty strings")
+    return command
+
+
 def main() -> int:
     parser = argparse.ArgumentParser()
     parser.add_argument("--container", required=True)
@@ -64,9 +71,11 @@ def main() -> int:
                 write_state(state_path, state)
                 return 0
             if args.completion_command_json:
-                command = json.loads(args.completion_command_json.read_text(encoding="utf-8"))
-                if not isinstance(command, list) or not command or not all(isinstance(x, str) and x for x in command):
+                try:
+                    command = load_completion_command(args.completion_command_json)
+                except (OSError, json.JSONDecodeError, ValueError) as exc:
                     state["status"] = "invalid_completion_command"
+                    state["completion_handoff_error"] = str(exc)
                     write_state(state_path, state)
                     return 4
                 result = subprocess.run(command, check=False)