diff --git a/src/documents/pdf_ops.py b/src/documents/pdf_ops.py
new file mode 100644
index 000000000..a8abbcd16
--- /dev/null
+++ b/src/documents/pdf_ops.py
@@ -0,0 +1,180 @@
+"""
+Pure PDF page operations used by documents.bulk_edit.
+
+This module deliberately knows nothing about Django, Celery or the documents
+app: callers resolve documents, choose output paths and queue work. Every
+function that writes a PDF removes unreferenced resources before saving.
+
+pikepdf is always called as ``pikepdf.open(...)`` / ``pikepdf.new()`` (never
+``from pikepdf import open``) so tests can patch those module attributes.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+from typing import NamedTuple
+
+import pikepdf
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+    from collections.abc import Iterable
+    from collections.abc import Mapping
+    from collections.abc import Sequence
+    from pathlib import Path
+    from types import TracebackType
+
+
+class PageSpec(NamedTuple):
+    """One page of an output PDF: a 1-indexed source page, optionally rotated."""
+
+    page: int
+    rotate: int = 0  # relative degrees, 0 leaves the page alone
+
+
+def _require_positive(pages: Iterable[int]) -> None:
+    for page in pages:
+        if page < 1:
+            raise ValueError(f"Page numbers start at 1, got {page}")
+
+
+def rotate_pdf(src: Path, dst: Path, degrees: int) -> None:
+    """Rotate every page relatively, in place, keeping document-level data."""
+    with pikepdf.open(src) as pdf:
+        for page in pdf.pages:
+            page.rotate(degrees, relative=True)
+        pdf.remove_unreferenced_resources()
+        pdf.save(dst)
+
+
+def remove_pages(src: Path, dst: Path, pages: Iterable[int]) -> None:
+    """
+    Remove 1-indexed pages in place, keeping document-level data.
+
+    Duplicates are ignored. Pages are removed highest first so earlier removals
+    never shift the index of later ones.
+    """
+    unique = sorted(set(pages))
+    _require_positive(unique)
+    with pikepdf.open(src) as pdf:
+        for page_num in reversed(unique):
+            del pdf.pages[page_num - 1]
+        pdf.remove_unreferenced_resources()
+        pdf.save(dst)
+
+
+def build_pdfs(
+    src: Path,
+    outputs: Sequence[tuple[Sequence[PageSpec], Callable[[], Path]]],
+) -> list[Path]:
+    """
+    Build one new PDF per output from pages of ``src``, opening ``src`` once.
+
+    Each output is ``(page_specs, make_dst)``. ``make_dst`` is called after that
+    output's pages are copied and immediately before it is saved, so a bad page
+    number never leaves a destination behind. Document-level data (Info, XMP,
+    outlines) is not carried over. Returns the written paths in output order.
+    """
+    for specs, _ in outputs:
+        _require_positive(spec.page for spec in specs)
+
+    written: list[Path] = []
+    with pikepdf.open(src) as source:
+        for specs, make_dst in outputs:
+            dst = pikepdf.new()
+            for spec in specs:
+                dst.pages.append(source.pages[spec.page - 1])
+                if spec.rotate:
+                    dst.pages[-1].rotate(spec.rotate, relative=True)
+            dst.remove_unreferenced_resources()
+            path = make_dst()
+            dst.save(path)
+            dst.close()
+            written.append(path)
+    return written
+
+
+def validate_page_operations(
+    operations: Sequence[Mapping[str, int]],
+    *,
+    single_output: bool,
+) -> int:
+    """
+    Validate ``edit_pdf`` style operations and return the output document count.
+
+    Each operation has ``page`` and optionally ``rotate`` and ``doc`` (the output
+    document index, default 0). The bounds rule is kept as it was: a ``doc`` index
+    must be below the number of operations.
+    """
+    if not operations:
+        raise ValueError("Output document index is out of bounds")
+
+    max_idx = max(op.get("doc", 0) for op in operations)
+    if single_output and max_idx > 0:
+        raise ValueError("Multiple output documents specified")
+
+    if any(
+        op.get("doc", 0) < 0 or op.get("doc", 0) >= len(operations) for op in operations
+    ):
+        raise ValueError("Output document index is out of bounds")
+
+    return max_idx + 1
+
+
+def needs_decrypt(src: Path) -> bool:
+    """
+    True if ``src`` is encrypted. A PDF that needs a password to open at all
+    counts as encrypted.
+    """
+    try:
+        with pikepdf.open(src) as pdf:
+            return bool(pdf.is_encrypted)
+    except pikepdf.PasswordError:
+        return True
+
+
+def decrypt_pdf(src: Path, make_dst: Callable[[], Path], password: str) -> Path:
+    """
+    Write an unencrypted copy of ``src`` and return its path.
+
+    ``make_dst`` is only called once the password has been accepted, so a wrong
+    password never leaves a destination behind.
+    """
+    with pikepdf.open(src, password=password) as pdf:
+        pdf.remove_unreferenced_resources()
+        dst = make_dst()
+        pdf.save(dst)
+    return dst
+
+
+class PdfMerger:
+    """
+    Accumulates the pages of several PDFs into one new PDF.
+
+    ``add`` raises if a source cannot be read; deciding whether to skip it is the
+    caller's policy. Use as a context manager so the merged PDF is closed.
+    """
+
+    def __init__(self) -> None:
+        self._pdf = pikepdf.new()
+        self._version: str = self._pdf.pdf_version
+
+    def __enter__(self) -> PdfMerger:
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc: BaseException | None,
+        tb: TracebackType | None,
+    ) -> None:
+        self._pdf.close()
+
+    def add(self, path: Path) -> None:
+        with pikepdf.open(str(path)) as pdf:
+            self._version = max(self._version, pdf.pdf_version)
+            self._pdf.pages.extend(pdf.pages)
+
+    def save(self, dst: Path) -> None:
+        self._pdf.remove_unreferenced_resources()
+        self._pdf.save(dst, min_version=self._version)
diff --git a/src/documents/tests/test_pdf_ops.py b/src/documents/tests/test_pdf_ops.py
new file mode 100644
index 000000000..747d20946
--- /dev/null
+++ b/src/documents/tests/test_pdf_ops.py
@@ -0,0 +1,385 @@
+"""
+Tests for documents.pdf_ops.
+
+These use real PDFs from the sample directories. No database, Celery or mocks.
+Pages are compared by a hash of their content stream, so page identity and order
+are easy to assert.
+"""
+
+import ast
+import hashlib
+from collections.abc import Callable
+from pathlib import Path
+
+import pikepdf
+import pytest
+
+from documents import pdf_ops
+from documents.pdf_ops import PageSpec
+
+SRC_ROOT = Path(__file__).parents[2]
+SAMPLES = Path(__file__).parent / "samples"
+THREE_PAGES = SAMPLES / "documents" / "originals" / "0000002.pdf"
+TWELVE_PAGES = SAMPLES / "barcodes" / "split-by-asn-2.pdf"
+ENCRYPTED = SAMPLES / "password-is-test.pdf"
+SIGNED = SRC_ROOT / "paperless" / "tests" / "samples" / "tesseract" / "signed.pdf"
+
+
+def _page_fingerprint(page: pikepdf.Page) -> str:
+    contents = page.obj.get("/Contents")
+    streams = list(contents) if isinstance(contents, pikepdf.Array) else [contents]
+    return hashlib.sha1(b"".join(s.read_bytes() for s in streams)).hexdigest()
+
+
+def fingerprints(path: Path) -> list[str]:
+    with pikepdf.open(path) as pdf:
+        return [_page_fingerprint(page) for page in pdf.pages]
+
+
+def rotations(path: Path) -> list[int]:
+    with pikepdf.open(path) as pdf:
+        return [int(page.obj.get("/Rotate", 0)) for page in pdf.pages]
+
+
+def docinfo_keys(path: Path) -> set[str]:
+    with pikepdf.open(path) as pdf:
+        return set(pdf.docinfo.keys())
+
+
+def constant(path: Path) -> Callable[[], Path]:
+    return lambda: path
+
+
+@pytest.fixture
+def source_fingerprints() -> list[str]:
+    fps = fingerprints(THREE_PAGES)
+    assert len(set(fps)) == 3, "sample must have three distinct pages"
+    return fps
+
+
+class TestRotatePdf:
+    def test_rotation_is_relative_and_applies_to_every_page(self, tmp_path: Path):
+        once = tmp_path / "once.pdf"
+        twice = tmp_path / "twice.pdf"
+
+        pdf_ops.rotate_pdf(THREE_PAGES, once, 90)
+        pdf_ops.rotate_pdf(once, twice, 90)
+
+        assert rotations(once) == [90, 90, 90]
+        assert rotations(twice) == [180, 180, 180]
+        assert fingerprints(twice) == fingerprints(THREE_PAGES)
+
+    def test_keeps_document_info(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.rotate_pdf(THREE_PAGES, dst, 90)
+
+        assert "/Creator" in docinfo_keys(dst)
+
+
+class TestRemovePages:
+    def test_removes_selected_pages(
+        self,
+        tmp_path: Path,
+        source_fingerprints: list[str],
+    ):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.remove_pages(THREE_PAGES, dst, [2])
+
+        assert fingerprints(dst) == [source_fingerprints[0], source_fingerprints[2]]
+
+    def test_duplicate_page_numbers_remove_the_page_once(
+        self,
+        tmp_path: Path,
+        source_fingerprints: list[str],
+    ):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.remove_pages(THREE_PAGES, dst, [2, 2])
+
+        assert fingerprints(dst) == [source_fingerprints[0], source_fingerprints[2]]
+
+    def test_unordered_pages(self, tmp_path: Path, source_fingerprints: list[str]):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.remove_pages(THREE_PAGES, dst, [3, 1])
+
+        assert fingerprints(dst) == [source_fingerprints[1]]
+
+    def test_empty_list_keeps_every_page(
+        self,
+        tmp_path: Path,
+        source_fingerprints: list[str],
+    ):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.remove_pages(THREE_PAGES, dst, [])
+
+        assert fingerprints(dst) == source_fingerprints
+
+    def test_removing_every_page_writes_an_empty_pdf(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.remove_pages(THREE_PAGES, dst, [1, 2, 3])
+
+        assert fingerprints(dst) == []
+
+    def test_keeps_document_info(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.remove_pages(THREE_PAGES, dst, [1])
+
+        assert "/Creator" in docinfo_keys(dst)
+
+    @pytest.mark.parametrize("bad_page", [0, -1])
+    def test_rejects_pages_below_one(self, tmp_path: Path, bad_page: int):
+        dst = tmp_path / "out.pdf"
+
+        with pytest.raises(ValueError, match="start at 1"):
+            pdf_ops.remove_pages(THREE_PAGES, dst, [1, bad_page])
+
+        assert not dst.exists()
+
+    def test_page_past_the_end_raises_and_writes_nothing(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+
+        with pytest.raises(IndexError):
+            pdf_ops.remove_pages(THREE_PAGES, dst, [99])
+
+        assert not dst.exists()
+
+
+class TestBuildPdfs:
+    def test_selects_and_orders_pages(
+        self,
+        tmp_path: Path,
+        source_fingerprints: list[str],
+    ):
+        dst = tmp_path / "out.pdf"
+
+        written = pdf_ops.build_pdfs(
+            THREE_PAGES,
+            [([PageSpec(3), PageSpec(1)], constant(dst))],
+        )
+
+        assert written == [dst]
+        assert fingerprints(dst) == [source_fingerprints[2], source_fingerprints[0]]
+
+    def test_rotates_only_the_requested_pages(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.build_pdfs(
+            THREE_PAGES,
+            [([PageSpec(1), PageSpec(2, 90), PageSpec(3, 180)], constant(dst))],
+        )
+
+        assert rotations(dst) == [0, 90, 180]
+
+    def test_writes_one_file_per_output_in_order(self, tmp_path: Path):
+        first = tmp_path / "first.pdf"
+        second = tmp_path / "second.pdf"
+        source = fingerprints(TWELVE_PAGES)
+
+        written = pdf_ops.build_pdfs(
+            TWELVE_PAGES,
+            [
+                ([PageSpec(p) for p in (1, 2, 3)], constant(first)),
+                ([PageSpec(p) for p in range(4, 13)], constant(second)),
+            ],
+        )
+
+        assert written == [first, second]
+        assert fingerprints(first) == source[:3]
+        assert fingerprints(second) == source[3:]
+
+    def test_empty_page_list_writes_a_zero_page_file(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+
+        pdf_ops.build_pdfs(THREE_PAGES, [([], constant(dst))])
+
+        assert fingerprints(dst) == []
+
+    def test_does_not_carry_over_document_info(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+        assert "/Creator" in docinfo_keys(THREE_PAGES)
+
+        pdf_ops.build_pdfs(THREE_PAGES, [([PageSpec(1)], constant(dst))])
+
+        assert "/Creator" not in docinfo_keys(dst)
+
+    def test_destination_is_not_requested_when_a_page_is_out_of_range(
+        self,
+        tmp_path: Path,
+    ):
+        requested: list[Path] = []
+
+        def make_dst() -> Path:
+            requested.append(tmp_path / "out.pdf")
+            return requested[-1]
+
+        with pytest.raises(IndexError):
+            pdf_ops.build_pdfs(THREE_PAGES, [([PageSpec(99)], make_dst)])
+
+        assert requested == []
+
+    @pytest.mark.parametrize("bad_page", [0, -1])
+    def test_rejects_pages_below_one_before_opening_anything(
+        self,
+        tmp_path: Path,
+        bad_page: int,
+    ):
+        requested: list[Path] = []
+
+        def make_dst() -> Path:
+            requested.append(tmp_path / "out.pdf")
+            return requested[-1]
+
+        with pytest.raises(ValueError, match="start at 1"):
+            pdf_ops.build_pdfs(
+                THREE_PAGES,
+                [([PageSpec(1)], make_dst), ([PageSpec(bad_page)], make_dst)],
+            )
+
+        assert requested == []
+
+
+class TestValidatePageOperations:
+    def test_returns_the_output_count(self):
+        operations = [{"page": 1}, {"page": 2}, {"page": 3}]
+
+        assert pdf_ops.validate_page_operations(operations, single_output=True) == 1
+
+    def test_gap_in_output_indices_counts_up_to_the_highest(self):
+        operations = [
+            {"page": 1, "doc": 0},
+            {"page": 2, "doc": 2},
+            {"page": 3, "doc": 0},
+        ]
+
+        count = pdf_ops.validate_page_operations(operations, single_output=False)
+
+        assert count == 3
+
+    def test_empty_operations_are_rejected(self):
+        with pytest.raises(ValueError, match="index is out of bounds"):
+            pdf_ops.validate_page_operations([], single_output=False)
+
+    def test_multiple_outputs_rejected_when_single_output_required(self):
+        operations = [{"page": 1, "doc": 0}, {"page": 2, "doc": 1}]
+
+        with pytest.raises(ValueError, match="Multiple output documents"):
+            pdf_ops.validate_page_operations(operations, single_output=True)
+
+    @pytest.mark.parametrize("doc", [-1, 2, 2**32])
+    def test_output_index_out_of_bounds(self, doc: int):
+        operations = [{"page": 1, "doc": 0}, {"page": 2, "doc": doc}]
+
+        with pytest.raises(ValueError, match="index is out of bounds"):
+            pdf_ops.validate_page_operations(operations, single_output=False)
+
+
+class TestDecrypt:
+    def test_needs_decrypt(self):
+        assert pdf_ops.needs_decrypt(ENCRYPTED) is True
+        assert pdf_ops.needs_decrypt(THREE_PAGES) is False
+
+    def test_pdf_that_opens_without_a_password_but_is_flagged_encrypted(self):
+        assert pdf_ops.needs_decrypt(SIGNED) is True
+
+    def test_decrypt_writes_an_unencrypted_copy(self, tmp_path: Path):
+        dst = tmp_path / "out.pdf"
+
+        result = pdf_ops.decrypt_pdf(ENCRYPTED, constant(dst), "test")
+
+        assert result == dst
+        assert pdf_ops.needs_decrypt(dst) is False
+
+    def test_wrong_password_raises_and_never_requests_a_destination(
+        self,
+        tmp_path: Path,
+    ):
+        requested: list[Path] = []
+
+        def make_dst() -> Path:
+            requested.append(tmp_path / "out.pdf")
+            return requested[-1]
+
+        with pytest.raises(pikepdf.PasswordError):
+            pdf_ops.decrypt_pdf(ENCRYPTED, make_dst, "wrong")
+
+        assert requested == []
+
+
+class TestPdfMerger:
+    def test_pages_are_appended_in_the_order_added(
+        self,
+        tmp_path: Path,
+        source_fingerprints: list[str],
+    ):
+        reordered = tmp_path / "reordered.pdf"
+        merged = tmp_path / "merged.pdf"
+        pdf_ops.build_pdfs(
+            THREE_PAGES,
+            [([PageSpec(3), PageSpec(1)], constant(reordered))],
+        )
+
+        with pdf_ops.PdfMerger() as merger:
+            merger.add(reordered)
+            merger.add(THREE_PAGES)
+            merger.save(merged)
+
+        assert fingerprints(merged) == [
+            source_fingerprints[2],
+            source_fingerprints[0],
+            *source_fingerprints,
+        ]
+
+    def test_output_version_is_at_least_the_highest_source_version(
+        self,
+        tmp_path: Path,
+    ):
+        merged = tmp_path / "merged.pdf"
+        with pikepdf.open(TWELVE_PAGES) as pdf:
+            source_versions = [pdf.pdf_version]
+        with pikepdf.open(THREE_PAGES) as pdf:
+            source_versions.append(pdf.pdf_version)
+
+        with pdf_ops.PdfMerger() as merger:
+            merger.add(TWELVE_PAGES)
+            merger.add(THREE_PAGES)
+            merger.save(merged)
+
+        with pikepdf.open(merged) as pdf:
+            assert pdf.pdf_version >= max(source_versions)
+
+    def test_unreadable_source_raises_so_the_caller_can_skip_it(
+        self,
+        tmp_path: Path,
+    ):
+        garbage = tmp_path / "garbage.pdf"
+        garbage.write_bytes(b"not a pdf")
+
+        with pdf_ops.PdfMerger() as merger:
+            with pytest.raises(pikepdf.PdfError):
+                merger.add(garbage)
+
+
+def test_pdf_ops_imports_only_the_standard_library_and_pikepdf():
+    tree = ast.parse(Path(pdf_ops.__file__).read_text())
+    imported: set[str] = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            imported.update(alias.name.split(".")[0] for alias in node.names)
+        elif isinstance(node, ast.ImportFrom) and node.level == 0 and node.module:
+            imported.add(node.module.split(".")[0])
+
+    coupled = imported & {
+        "django",
+        "celery",
+        "documents",
+        "paperless",
+        "paperless_mail",
+        "paperless_ai",
+    }
+    assert not coupled