diff --git a/brukerapi/dataset.py b/brukerapi/dataset.py
index 2df6dc1..8645782 100644
--- a/brukerapi/dataset.py
+++ b/brukerapi/dataset.py
@@ -1577,9 +1577,10 @@ def acquisition_affines(self):
         the reconstruction's ``VisuFGOrderDesc`` records; for ``NI == NSLICES``
         object *k* is slice *k*.
 
-        :raise: :UnsupportedDatasetType: for anything but a 2-D or 3-D spatial
+        :raises UnsupportedDatasetType: for anything but a 2-D or 3-D spatial
             fid/rawdata acquisition, one without a gradient matrix, or a PV360
             one without a known subject position
+
         """
         if self.type not in ("fid", "rawdata"):
             raise UnsupportedDatasetType(f"an acquisition affine for {self.path}, which is not a raw acquisition,")
diff --git a/docs/source/compatibility.rst b/docs/source/compatibility.rst
index 16e5c19..091b53a 100644
--- a/docs/source/compatibility.rst
+++ b/docs/source/compatibility.rst
@@ -113,3 +113,12 @@ Data contract and limitations
   ``VisuCoreFrameThickness`` stays the source for the thickness (for a 3-D
   acquisition that parameter is the whole slab, not a plane step). It is
   available wherever ``affine_of_package`` is.
+* ``Dataset.acquisition_affines()`` generates the acquisition-side equivalent
+  for 2-D/3-D spatial FID and rawdata datasets, one transform per 2-D slice or
+  one per 3-D volume. It uses the encoded matrix size rather than the raw sample
+  array shape and returns the same Visu/DICOM patient frame as ``affine``.
+  ParaVision 5.1/6/7 gradient matrices already encode the subject position;
+  PV360 matrices require the declared ``ACQ_patient_pos`` conversion. Method
+  slice-package geometry is a warning-emitting fallback for missing ``acqp``
+  values. Spectroscopy/CSI and incomplete or unknown geometry raise
+  ``UnsupportedDatasetType``.
diff --git a/docs/source/dataset.rst b/docs/source/dataset.rst
index dacc658..fc710de 100644
--- a/docs/source/dataset.rst
+++ b/docs/source/dataset.rst
@@ -36,6 +36,53 @@ Accessing the latter emits ``FutureWarning``; use ``raw`` or ``kspace`` for
 new code. EPI and non-Cartesian PV360 jobs are intentionally not reconstructed
 by ``kspace``.
 
+.. _raw-acquisition-geometry:
+
+Raw acquisition geometry
+------------------------
+
+Spatial FID and PV360 rawdata datasets can generate affine matrices before an
+image has been reconstructed. ``acquisition_affines()`` returns one matrix per
+2-D slice; ``acquisition_affine(index)`` returns a selected matrix. A 3-D
+acquisition has one matrix whose third column is the partition step.
+
+.. code-block:: python
+
+   acquisition = Dataset("path/to/fid")
+   affines = acquisition.acquisition_affines()
+   first_slice_affine = acquisition.acquisition_affine(0)
+
+Each 4x4 matrix maps a voxel index in the image produced by Fourier transforming
+the encoded k-space matrix to millimetres in the Visu/DICOM patient frame:
+
+.. code-block:: python
+
+   patient_position = first_slice_affine @ [i, j, 0, 1]
+
+The first two columns contain the read and phase voxel steps, the third contains
+the slice or partition step, and the last column locates voxel ``(0, 0, 0)``.
+This is the same patient frame used by a reconstructed 2dseq
+``Dataset.affine``, so acquisition and reconstruction geometry can be compared
+directly. To produce a NIfTI RAS affine, negate the patient x and y axes:
+
+.. code-block:: python
+
+   import numpy as np
+
+   nifti_affine = np.diag([-1, -1, 1, 1]) @ first_slice_affine
+
+Geometry is derived primarily from ``ACQ_grad_matrix``, the ``ACQ_*_offset``
+values and ``ACQ_fov``. Method slice-package parameters are used as fallbacks
+when acquisition parameters are absent, with a ``RuntimeWarning`` because they
+are not equivalent for every ParaVision version. Spectroscopy/CSI, non-spatial
+or non-2-D/3-D acquisitions, missing orientation metadata, and PV360 data with
+an unknown subject position raise ``UnsupportedDatasetType``.
+
+The returned list is indexed by physical slice, not by every acquisition
+object. If echoes, repetitions, or other dimensions make ``NI`` differ from
+``NSLICES``, their object-to-slice nesting can only be recovered from the
+reconstruction metadata.
+
 Metadata views
 --------------
 
diff --git a/docs/source/tutorials/how-to-fid.rst b/docs/source/tutorials/how-to-fid.rst
index 5d62735..43cbac9 100644
--- a/docs/source/tutorials/how-to-fid.rst
+++ b/docs/source/tutorials/how-to-fid.rst
@@ -58,6 +58,19 @@ metadata. An explicit override is available when inference is ambiguous:
 
    dataset = Dataset('path/to/fid', scheme_id='RADIAL')
 
+For a 2-D or 3-D spatial acquisition, generate the geometry of the image that
+a Fourier transform of the encoded k-space matrix would produce:
+
+.. code-block:: python
+
+   slice_affines = dataset.acquisition_affines()
+   first_slice_affine = dataset.acquisition_affine(0)
+
+The matrices map voxel indices to millimetres in the Visu/DICOM patient frame,
+the same frame as a reconstructed 2dseq ``affine``. There is one matrix per
+2-D slice and one for a 3-D volume. See :ref:`raw-acquisition-geometry` for the
+coordinate convention, limitations, and NIfTI conversion.
+
 Random-access ``mmap=True`` is currently supported for ``2dseq`` only. Load a
 FID normally, then select the desired k-space array slice.
 
diff --git a/docs/source/tutorials/how-to-rawdata.rst b/docs/source/tutorials/how-to-rawdata.rst
index 26db5f1..1c4b90b 100644
--- a/docs/source/tutorials/how-to-rawdata.rst
+++ b/docs/source/tutorials/how-to-rawdata.rst
@@ -27,6 +27,19 @@ metadata-reordered array. BART consumers can request its 16-axis layout:
    kspace = rawdata.kspace
    bart_kspace = rawdata.to_kspace(bart=True)
 
+Generate patient-space geometry for a spatial rawdata acquisition in the same
+way as for a FID:
+
+.. code-block:: python
+
+   slice_affines = rawdata.acquisition_affines()
+   first_slice_affine = rawdata.acquisition_affine(0)
+
+The declared ``ACQ_patient_pos`` is applied when converting PV360 gradient
+directions to the Visu/DICOM patient frame. An unknown position raises
+``UnsupportedDatasetType``. See :ref:`raw-acquisition-geometry` for the full
+coordinate convention and limitations.
+
 ``rawdata.data`` is retained for compatibility and exposes the historical
 decoded job layout. It emits ``FutureWarning`` because that layout differs from
 FID ``data``. Prefer ``raw`` or ``kspace`` in new applications. EPI and