diff --git a/docs/contract.md b/docs/contract.md
index d9bda27..5700bb2 100644
--- a/docs/contract.md
+++ b/docs/contract.md
@@ -580,7 +580,10 @@ Computed from `plane_histograms.csv` rows alone (N38).
 `image_summaries(records, percentile=(1.0, 99.8))` returns one `ImageSummary` per image, sorted
 by condition, biological repeat, technical repeat and path, pooling all Z and T planes of the
 image with the group statistics of §1.2.3 (inverse-CDF percentiles, exact for 8- and 16-bit
-histograms). Images with different numbers of channels are refused with `ValueError` (N43).
+histograms). Images with different numbers of channels are refused with `ValueError` (N43), and so is an image whose Z or T planes have different pixel
+types (`IncompatibleSummariesError`, a `ValueError`). An `ImageSummary` whose per-channel tuples
+have different lengths is refused when made, and `intensity_checks` refuses summaries with
+different numbers of channels or different percentile definitions, with the same error.
 `ImageSummary.to_row()` gives the columns below as a dictionary, in this order (numbers as
 floats, a blank as `None`, `Histogram_Exact_C<i>` as a bool), and `summary_columns(n)` their
 names for `n` channels, so the calling brick can show the rows as a table or write them; the
diff --git a/src/morphobricks_normalization/__init__.py b/src/morphobricks_normalization/__init__.py
index 5fd431e..8adfcff 100644
--- a/src/morphobricks_normalization/__init__.py
+++ b/src/morphobricks_normalization/__init__.py
@@ -39,7 +39,12 @@
 from .rgb import apply_rgb, rgb_percentile
 from .scopes import group_key, group_name, scope_groups
 from .settings import MethodNotAllowedError, NormalizationSettings, check_settings
-from .summaries import ImageSummary, image_summaries, summary_columns
+from .summaries import (
+    ImageSummary,
+    IncompatibleSummariesError,
+    image_summaries,
+    summary_columns,
+)
 
 __all__ = [
     "OUTPUT_FILENAME",
@@ -47,6 +52,7 @@
     "HistogramSettings",
     "HistogramTail",
     "ImageSummary",
+    "IncompatibleSummariesError",
     "IntensityCheckSettings",
     "Issue",
     "MethodNotAllowedError",
diff --git a/src/morphobricks_normalization/checks.py b/src/morphobricks_normalization/checks.py
index 1deefca..3f9d045 100644
--- a/src/morphobricks_normalization/checks.py
+++ b/src/morphobricks_normalization/checks.py
@@ -43,7 +43,8 @@
     Issue,
 )
 from .normalization import NORMAL_CONSISTENCY
-from .summaries import ImageSummary
+from .records import different_channel_counts
+from .summaries import ImageSummary, IncompatibleSummariesError
 
 _KEYS = ("library_version", "outlier_limit", "saturation_limit", "min_images")
 
@@ -178,6 +179,8 @@ def intensity_checks(
     """
     settings = IntensityCheckSettings() if settings is None else settings
     summaries = sorted(summaries, key=_order)
+    _check_comparable(summaries)
+    _check_same_percentiles(summaries)
     return [
         *_image_outliers(summaries, settings),
         *_repeat_shifts(summaries, settings),
@@ -383,6 +386,34 @@ def _compare(value: float, others: list[float], limit: float) -> tuple[float, fl
     return (reference, deviation, distance) if distance > limit else None
 
 
+def _check_comparable(summaries: list[ImageSummary]) -> None:
+    """Refuse summaries that do not describe the same channels (issue #56)."""
+    if len({s.channel_count for s in summaries}) > 1:
+        raise IncompatibleSummariesError(
+            different_channel_counts(
+                [(s.source_relative_path, s.channel_count) for s in summaries], "These images"
+            )
+        )
+
+
+def _check_same_percentiles(summaries: list[ImageSummary]) -> None:
+    """Refuse summaries computed with different percentile definitions (issue #55)."""
+    by_definition: dict[tuple[float, float], list[str]] = {}
+    for s in summaries:
+        by_definition.setdefault(tuple(s.percentile), []).append(s.source_relative_path)
+    if len(by_definition) > 1:
+        parts = []
+        for (low, high), names in sorted(by_definition.items()):
+            shown = ", ".join(names[:3]) + (f" and {len(names) - 3} more" if len(names) > 3 else "")
+            parts.append(f"percentiles ({low:g}, {high:g}): {shown}")
+        raise IncompatibleSummariesError(
+            f"These images have summaries with different percentile definitions "
+            f"({'; '.join(parts)}). Values at different percentiles are different measurements and "
+            "cannot be compared. Summarize every "
+            "image with one percentile pair, with image_summaries(records, percentile=...)."
+        )
+
+
 def _order(summary: ImageSummary) -> tuple[str, ...]:
     # The order of image_summaries: labels from the top down, then the path.
     return (
diff --git a/src/morphobricks_normalization/summaries.py b/src/morphobricks_normalization/summaries.py
index ec3b115..064ff77 100644
--- a/src/morphobricks_normalization/summaries.py
+++ b/src/morphobricks_normalization/summaries.py
@@ -49,6 +49,14 @@
 )
 
 
+class IncompatibleSummariesError(ValueError):
+    """Raised for image summaries that cannot be compared or pooled with each other.
+
+    Different numbers of channels, percentile definitions or pixel types make the
+    values measure different things; nothing is converted or guessed.
+    """
+
+
 @dataclass(frozen=True)
 class ImageSummary:
     """Intensity summary of one image, all Z and T planes pooled per channel.
@@ -86,6 +94,21 @@ class ImageSummary:
     saturated_fraction: tuple[float | None, ...]
     exact: tuple[bool | None, ...]
 
+    def __post_init__(self) -> None:
+        """Refuse per-channel tuples of different lengths, or none (issue #56)."""
+        lengths = {
+            name: len(getattr(self, name))
+            for name in ("median", "p_low", "p_high", "saturated_fraction", "exact")
+        }
+        if len(set(lengths.values())) > 1 or lengths["median"] == 0:
+            found = ", ".join(f"{name}: {n}" for name, n in lengths.items())
+            raise IncompatibleSummariesError(
+                f"Image {self.source_relative_path}: its per-channel values do not have one "
+                f"entry per channel ({found}), so the channels of the image are not defined. "
+                "Give median, p_low, p_high, saturated_fraction and exact the same, non-zero "
+                "length, or build the summary with image_summaries()."
+            )
+
     @property
     def image_file(self) -> str:
         """File name of the image."""
@@ -161,6 +184,8 @@ def image_summaries(
     ValueError
         For invalid percentiles, rows that repeat a ``plane_id``, or images with
         different numbers of channels (N43).
+    IncompatibleSummariesError
+        For an image whose Z or T planes have different pixel types (issue #54).
     """
     low, high = (float(q) for q in percentile)
     if not 0 <= low < high <= 100:
@@ -179,7 +204,24 @@ def image_summaries(
     return [_summary(planes, (low, high)) for _, planes in sorted(images.items())]
 
 
+def _check_one_dtype(planes: list[PlaneRecord]) -> None:
+    """Refuse an image whose Z/T planes have different pixel types (issue #54)."""
+    by_dtype: dict[str, list[str]] = {}
+    for plane in sorted(planes, key=lambda r: (r.t, r.z)):
+        by_dtype.setdefault(str(plane.histograms[0].dtype), []).append(plane.plane_id)
+    if len(by_dtype) > 1:
+        parts = "; ".join(f"{dtype}: {', '.join(ids)}" for dtype, ids in sorted(by_dtype.items()))
+        raise IncompatibleSummariesError(
+            f"Image {planes[0].source_relative_path} has planes of different pixel types "
+            f"({parts}). The saturated fraction and the pooled statistics need one pixel type "
+            "per image, and the pixel types are not converted. Save the planes of the image "
+            "with one pixel type, then run the preprocessing brick again."
+        )
+
+
 def _summary(planes: list[PlaneRecord], percentile: tuple[float, float]) -> ImageSummary:
+    """Pool the planes of one image, channel by channel."""
+    _check_one_dtype(planes)
     first = planes[0]
     median, p_low, p_high, saturated, exact = [], [], [], [], []
     for channel in range(first.channel_count):
diff --git a/tests/test_checks.py b/tests/test_checks.py
index d352252..528d34f 100644
--- a/tests/test_checks.py
+++ b/tests/test_checks.py
@@ -8,6 +8,7 @@
 from morphobricks_normalization import (
     SINGLE,
     ImageSummary,
+    IncompatibleSummariesError,
     IntensityCheckSettings,
     Issue,
     PlaneRecord,
@@ -359,3 +360,54 @@ def test_other_warnings_have_no_channel():
     issue = Issue("warning", "repeat_missing_conditions", "message")
     assert issue.channel is None
     assert replace(issue, channel=1) != issue
+
+
+def _direct(path_, channels, percentile=(1.0, 99.8)):
+    """A directly built summary with ``channels`` channels."""
+    values = tuple(float(10 + c) for c in range(channels))
+    return ImageSummary(
+        source_relative_path=path_,
+        condition="c",
+        biological_repeat="r",
+        technical_repeat="",
+        dtype=np.dtype(np.uint8),
+        percentile=percentile,
+        median=values,
+        p_low=values,
+        p_high=tuple(v + 5 for v in values),
+        saturated_fraction=(0.0,) * channels,
+        exact=(True,) * channels,
+    )
+
+
+def test_summary_with_mismatched_channel_tuples_is_refused():
+    """Issue #56: per-channel tuples of different lengths name the image."""
+    good = _direct("a.tif", 2)
+    with pytest.raises(IncompatibleSummariesError, match=r"a\.tif.*p_high: 1"):
+        replace(good, p_high=(1.0,))
+    with pytest.raises(IncompatibleSummariesError, match=r"a\.tif"):
+        replace(good, median=(), p_low=(), p_high=(), saturated_fraction=(), exact=())
+
+
+@pytest.mark.parametrize("order", [(1, 2), (2, 1)])
+def test_summaries_with_different_channel_counts_are_refused(order):
+    """Issue #56: one- and two-channel images in either order, naming both images."""
+    summaries = [_direct(f"img_{i}.tif", n) for i, n in enumerate(order * 3)]
+    with pytest.raises(IncompatibleSummariesError, match=r"1 channel\(s\).*2 channel\(s\)"):
+        intensity_checks(summaries)
+
+
+@pytest.mark.parametrize("odd_first", [True, False])
+@pytest.mark.parametrize("odd", [(1.0, 95.0), (5.0, 99.8)])
+def test_summaries_with_different_percentiles_are_refused(odd, odd_first):
+    """Issue #55: a different high or low percentile is refused, in either order."""
+    same = [_direct(f"img_{i}.tif", 1) for i in range(5)]
+    other = _direct("odd.tif", 1, percentile=odd)
+    summaries = [other, *same] if odd_first else [*same, other]
+    with pytest.raises(IncompatibleSummariesError, match=r"percentile definitions.*odd\.tif"):
+        intensity_checks(summaries)
+
+
+def test_summaries_with_the_same_percentiles_are_accepted():
+    summaries = [_direct(f"img_{i}.tif", 1, percentile=(5.0, 95.0)) for i in range(5)]
+    assert intensity_checks(summaries) == []
diff --git a/tests/test_summaries.py b/tests/test_summaries.py
index 477019b..80ff6e8 100644
--- a/tests/test_summaries.py
+++ b/tests/test_summaries.py
@@ -7,6 +7,7 @@
 
 from morphobricks_normalization import (
     SINGLE,
+    IncompatibleSummariesError,
     PlaneRecord,
     compute_plane_histogram,
     image_summaries,
@@ -217,3 +218,30 @@ def test_summaries_from_the_written_table_equal_those_from_the_records(tmp_path)
 
 def test_no_records_no_summaries():
     assert image_summaries([]) == []
+
+
+@pytest.mark.parametrize("first_is_small", [True, False])
+@pytest.mark.parametrize(
+    ("small", "large"), [(np.uint8, np.uint16), (np.uint16, np.int16), (np.uint8, np.int8)]
+)
+def test_planes_of_one_image_with_different_pixel_types_are_refused(small, large, first_is_small):
+    """Issue #54: no pixel is at either type's maximum, yet the first dtype used to decide."""
+    a = record("a.tif", [np.full((8, 8), 100, dtype=small)], z=0)
+    b = record("a.tif", [np.full((8, 8), 100, dtype=large)], z=1)
+    with pytest.raises(IncompatibleSummariesError, match=r"a\.tif.*different pixel types") as info:
+        image_summaries([a, b] if first_is_small else [b, a])
+    assert str(np.dtype(small)) in str(info.value) and str(np.dtype(large)) in str(info.value)
+
+
+def test_same_pixel_type_with_varying_planes_is_accepted():
+    records = [record("a.tif", [np.full((8, 8), 10 * (z + 1), np.uint16)], z=z) for z in range(3)]
+    (summary,) = image_summaries(records)
+    assert summary.saturated_fraction == (0.0,)
+
+
+def test_different_images_may_differ_in_pixel_type():
+    records = [
+        record("a.tif", [np.full((8, 8), 5, np.uint8)]),
+        record("b.tif", [np.full((8, 8), 5, np.uint16)]),
+    ]
+    assert [s.dtype for s in image_summaries(records)] == [np.uint8, np.uint16]
diff --git a/tools/code_rules_baseline.json b/tools/code_rules_baseline.json
index b19c9a6..997d7e3 100644
--- a/tools/code_rules_baseline.json
+++ b/tools/code_rules_baseline.json
@@ -40,7 +40,6 @@
     "src/morphobricks_normalization/settings.py _positive_number: # Ruff findings any-type": 1,
     "src/morphobricks_normalization/summaries.py _saturated_fraction: # Ruff findings missing-type-function-argument": 1,
     "src/morphobricks_normalization/summaries.py _saturated_fraction: private function of # lines without a docstring": 16,
-    "src/morphobricks_normalization/summaries.py _summary: private function of # lines without a docstring": 28,
     "tests/data/make_v#_fixture.py <module>: # Ruff findings module-import-not-at-top-of-file": 3
   }
 }