diff --git a/graphify/detect.py b/graphify/detect.py
index f76a4259f..32fdead47 100644
--- a/graphify/detect.py
+++ b/graphify/detect.py
@@ -1644,7 +1644,8 @@ def _on_walk_error(err: OSError) -> None:
 
     all_files.sort(key=lambda p: str(p))
 
-    converted_dir = root / GRAPHIFY_OUT / "converted"
+    out_base = Path(cache_root).resolve() if cache_root is not None else root
+    converted_dir = out_base / GRAPHIFY_OUT / "converted"
 
     for p in all_files:
         # For memory dir files, skip hidden/noise filtering
diff --git a/tests/test_detect.py b/tests/test_detect.py
index 972955297..36e6a4811 100644
--- a/tests/test_detect.py
+++ b/tests/test_detect.py
@@ -2557,6 +2557,49 @@ def test_convert_office_file_outside_root_falls_back(tmp_path, monkeypatch):
     assert out1 is not None and out1.name == out2.name
 
 
+def test_detect_office_conversion_respects_cache_root(tmp_path, monkeypatch):
+    """#2787: detect() with cache_root must write converted sidecars under
+    cache_root/GRAPHIFY_OUT/converted, leaving the scanned corpus untouched, while
+    keeping the sidecar filename hash anchored to the scan root."""
+    monkeypatch.setattr(detect_mod, "docx_to_markdown", lambda p: "# Spec\nConverted specification text.")
+
+    corpus = tmp_path / "corpus"
+    corpus.mkdir(parents=True)
+    cache_out = tmp_path / "cache_out"
+    cache_out.mkdir(parents=True)
+
+    doc_path = corpus / "spec.docx"
+    doc_path.write_bytes(b"placeholder")
+
+    result = detect(corpus, cache_root=cache_out)
+
+    # 1. Scanned corpus tree must not be mutated
+    assert not (corpus / detect_mod.GRAPHIFY_OUT).exists(), (
+        "detect() must not write graphify-out into the scanned corpus tree when cache_root is provided (#2787)"
+    )
+
+    # 2. Converted sidecar must exist under cache_root
+    converted_dir = cache_out / detect_mod.GRAPHIFY_OUT / "converted"
+    assert converted_dir.is_dir(), "converted directory must be created under cache_root"
+
+    # 3. Detection result must point to the redirected sidecar
+    doc_files = result["files"]["document"]
+    assert len(doc_files) == 1
+    sidecar_path = Path(doc_files[0])
+    assert sidecar_path.is_file()
+    assert sidecar_path.parent.resolve() == converted_dir.resolve()
+
+    # 4. Content must match expected conversion
+    content = sidecar_path.read_text(encoding="utf-8")
+    assert "<!-- converted from spec.docx -->" in content
+    assert "# Spec" in content
+
+    # 5. Sidecar filename must remain anchored to the corpus scan root
+    import hashlib
+    expected_hash = hashlib.sha256(unicodedata.normalize("NFC", "spec.docx").encode()).hexdigest()[:8]
+    assert sidecar_path.name == f"spec_{expected_hash}.md"
+
+
 def test_detect_keeps_env_source_dirs(tmp_path):
     """#2058: a real source directory named env/ or *_env/ with no virtualenv
     markers must be indexed, not silently pruned as a false-positive venv."""