diff --git a/src/vak/eval/frame_classification.py b/src/vak/eval/frame_classification.py
index 665a0bb4f..f14eae17e 100644
--- a/src/vak/eval/frame_classification.py
+++ b/src/vak/eval/frame_classification.py
@@ -1,5 +1,4 @@
 """Function that evaluates trained models in the frame classification family."""
-
 from __future__ import annotations
 
 import json
@@ -30,7 +29,7 @@ def eval_frame_classification_model(
     num_workers: int,
     frames_standardizer_path: str | pathlib.Path = None,
     post_tfm_kwargs: dict | None = None,
-) -> None:
+) -> dict[str, torch.Tensor]:
     """Evaluate a trained model.
 
     Parameters
@@ -215,10 +214,11 @@ def eval_frame_classification_model(
         logger=trainer_logger,
     )
     # TODO: check for hasattr(model, test_step) and if so run test
+
     # below, [0] because validate returns list of dicts, length of no. of val loaders
     metric_vals = trainer.validate(model, dataloaders=val_loader)[0]
-    metric_vals = {f"avg_{k}": v for k, v in metric_vals.items()}
-    for metric_name, metric_val in metric_vals.items():
+    metric_vals_for_df = {f"avg_{k}": v for k, v in metric_vals.items()}
+    for metric_name, metric_val in metric_vals_for_df.items():
         if metric_name.startswith("avg_"):
             logger.info(f"{metric_name}: {metric_val:0.5f}")
 
@@ -237,7 +237,7 @@ def eval_frame_classification_model(
     # order metrics by name to be extra sure they will be consistent across runs
     row.update(
         sorted(
-            [(k, v) for k, v in metric_vals.items() if k.startswith("avg_")]
+            [(k, v) for k, v in metric_vals_for_df.items() if k.startswith("avg_")]
         )
     )
 
@@ -249,3 +249,6 @@ def eval_frame_classification_model(
     eval_df.to_csv(
         eval_csv_path, index=False
     )  # index is False to avoid having "Unnamed: 0" column when loading
+
+    return metric_vals
+
diff --git a/src/vak/transforms/frame_labels/functional.py b/src/vak/transforms/frame_labels/functional.py
index e3b89bca5..0c05a68dd 100644
--- a/src/vak/transforms/frame_labels/functional.py
+++ b/src/vak/transforms/frame_labels/functional.py
@@ -556,6 +556,15 @@ def postprocess(
     frame_labels : numpy.ndarray
         Vector of frame labels after post-processing is applied.
     """
+    # NOTE: we need to clone to avoid mutating tensors inside a model,
+    # since calling `torch.tensor.numpy()` inside the model gives us 
+    # a numpy array that shares memory with the tensor.
+    # Not doing so can make metrics wrong, 
+    # e.g., for frame labels *before* post-processing, 
+    # if we mutate that tensor here.
+    # see:https://github.com/vocalpy/vak/issues/821
+    frame_labels = np.copy(frame_labels)
+
     frame_labels = row_or_1d(frame_labels)
     if boundary_labels is not None:
         boundary_labels = row_or_1d(boundary_labels)
diff --git a/tests/test_eval/test_frame_classification.py b/tests/test_eval/test_frame_classification.py
index 8ac65539a..3c58cade9 100644
--- a/tests/test_eval/test_frame_classification.py
+++ b/tests/test_eval/test_frame_classification.py
@@ -2,13 +2,11 @@
 import pytest
 
 import vak.config
-import vak.common.constants
-import vak.common.paths
 import vak.eval.frame_classification
 
 
 # written as separate function so we can re-use in tests/unit/test_cli/test_eval.py
-def assert_eval_output_matches_expected(model_name, output_dir):
+def assert_eval_saves_one_csv(model_name, output_dir):
     eval_csv = sorted(output_dir.glob(f"eval_{model_name}*csv"))
     assert len(eval_csv) == 1
 
@@ -22,9 +20,13 @@ def assert_eval_output_matches_expected(model_name, output_dir):
     # use ToLabelsWithPostprocessing with *just* majority_vote
     {'majority_vote': True, 'min_segment_dur': None},
     # use ToLabelsWithPostprocessing with *just* min_segment_dur
-    {'majority_vote': False, 'min_segment_dur': 0.002},
+    # NOTE the value of `min_segment_dur=0.05` will make metrics after post-processing *worse*
+    # but we do that to make sure that this parameter **alone** is having some effect.
+    # For our current test data, the "good" value of 0.01 we used can in some cases 
+    # have no effect; we want to be sure that's not due to some bug
+    {'majority_vote': False, 'min_segment_dur': 0.05},
     # use ToLabelsWithPostprocessing with majority_vote *and* min_segment_dur
-    {'majority_vote': True, 'min_segment_dur': 0.002},
+    {'majority_vote': True, 'min_segment_dur': 0.05},
 ]
 
 
@@ -69,7 +71,7 @@ def test_eval_frame_classification_model(
     )
     cfg = vak.config.Config.from_toml_path(toml_path)
 
-    vak.eval.frame_classification.eval_frame_classification_model(
+    metric_vals: dict = vak.eval.frame_classification.eval_frame_classification_model(
         model_config=cfg.eval.model.asdict(),
         dataset_config=cfg.eval.dataset.asdict(),
         trainer_config=cfg.eval.trainer.asdict(),
@@ -81,7 +83,34 @@ def test_eval_frame_classification_model(
         post_tfm_kwargs=post_tfm_kwargs,
     )
 
-    assert_eval_output_matches_expected(cfg.eval.model.name, output_dir)
+    assert_eval_saves_one_csv(cfg.eval.model.name, output_dir)
+
+    # REGRESSION TEST: these assertions test that
+    # we fixed https://github.com/vocalpy/vak/issues/821
+    # Long-term this may not be the right place to test this, 
+    # but it will work to get out a quick fix;
+    # NOTE also this is fragile -- I expect the transforms to change the metrics 
+    # for this exact data, but of course you could have a case where 
+    # neither `majority_vote` nor `min_segment_dur` have any effect 
+    # even when `majority_vote=True` and `min_segment_dur > 0.0
+    if post_tfm_kwargs is not None:
+        if post_tfm_kwargs["majority_vote"] is True or (
+            post_tfm_kwargs["min_segment_dur"] is not None and post_tfm_kwargs["min_segment_dur"] > 0.0
+        ):
+            for metric_name in (
+                "acc", "levenshtein", "character_error_rate",
+            ):
+                wout_tfm = metric_vals[f"val_{metric_name}_epoch"]
+                with_tfm = metric_vals[f"val_{metric_name}_tfm_epoch"]
+                assert wout_tfm != with_tfm
+            # for IR metrics, sometimes recall is unchanged
+            # we mainly want to be sure they are not *all* literally the sma value
+            assert not all(
+                [
+                    metric_vals[f"val_{ir_metric_name}_epoch"] == metric_vals[f"val_{ir_metric_name}_tfm_epoch"]
+                    for ir_metric_name in ("precision", "recall", "fscore", "rval")
+                ]
+            )
 
 
 @pytest.mark.parametrize(
diff --git a/tests/test_eval/test_parametric_umap.py b/tests/test_eval/test_parametric_umap.py
index b1744291d..d37cafa9e 100644
--- a/tests/test_eval/test_parametric_umap.py
+++ b/tests/test_eval/test_parametric_umap.py
@@ -7,7 +7,7 @@
 import vak.eval.frame_classification
 
 
-def assert_eval_output_matches_expected(model_name, output_dir):
+def assert_eval_saves_one_csv(model_name, output_dir):
     eval_csv = sorted(output_dir.glob(f"eval_{model_name}*csv"))
     assert len(eval_csv) == 1
 
@@ -57,7 +57,7 @@ def test_eval_parametric_umap_model(
         num_workers=cfg.eval.num_workers,
     )
 
-    assert_eval_output_matches_expected(cfg.eval.model.name, output_dir)
+    assert_eval_saves_one_csv(cfg.eval.model.name, output_dir)
 
 
 @pytest.mark.parametrize(