diff --git a/src/benchflow/eval_plan.py b/src/benchflow/eval_plan.py
index f2cd6b39c..af5fd9f44 100644
--- a/src/benchflow/eval_plan.py
+++ b/src/benchflow/eval_plan.py
@@ -264,6 +264,14 @@ def build_eval_plan(request: EvalCreateRequest) -> EvalPlan:
         raise EvalPlanError("--matrix currently requires --tasks-dir")
     if request.trials < 1:
         raise EvalPlanError("--trials must be >= 1")
+    if request.trials > 1 and request.matrix is None:
+        # Only the matrix expansion consumes trials; a plain run would silently
+        # do one trial per task while the caller believes it ran N. Per-trial
+        # requests created by run_matrix_eval bypass this planner by design.
+        raise EvalPlanError(
+            "--trials > 1 requires --matrix; to repeat one model, use a "
+            'single-entry matrix such as "models: {default: <model>}"'
+        )
     if request.expected_tasks is not None and request.expected_tasks < 1:
         raise EvalPlanError("--expected-tasks must be >= 1")
     if request.canonicalize not in {"none", "one-healthy-per-task"}:
diff --git a/tests/test_cli_arg_validation.py b/tests/test_cli_arg_validation.py
index 44cf1d382..e8d32128b 100644
--- a/tests/test_cli_arg_validation.py
+++ b/tests/test_cli_arg_validation.py
@@ -65,6 +65,14 @@ def test_build_concurrency_zero_rejected(tmp_path: Path):
     assert "--build-concurrency must be >= 1" in result.stderr
 
 
+def test_trials_without_matrix_rejected(tmp_path: Path):
+    """Guards PR #1064: plain runs must not silently ignore --trials."""
+    result = _invoke(tmp_path, "--sandbox", "docker", "--trials", "3")
+    assert result.exit_code == 1
+    assert "--trials > 1 requires --matrix" in result.stderr
+    assert "single-entry matrix" in result.stderr
+
+
 def test_skill_mode_bogus_clean_error(tmp_path: Path):
     result = _invoke(tmp_path, "--sandbox", "docker", "--skill-mode", "bogus")
     assert result.exit_code == 1