diff --git a/paperforge/worker/ocr_document.py b/paperforge/worker/ocr_document.py
index e348f2ce..2df91921 100644
--- a/paperforge/worker/ocr_document.py
+++ b/paperforge/worker/ocr_document.py
@@ -1797,6 +1797,16 @@ def _normalize_backmatter_roles_after_boundary(
     if tail_boundary is None or tail_boundary.spread_start is None or tail_boundary.spread_end is None:
         return
 
+    _POST_REF_PRESERVE_ROLES = {
+        "figure_inner_text",
+        "figure_caption",
+        "figure_caption_candidate",
+        "table_caption",
+        "table_caption_candidate",
+        "media_asset",
+        "table_html",
+    }
+
     tail_start = _effective_tail_start(tail_boundary, blocks)
     boundary_seen = False
     backmatter_started = False
@@ -1804,6 +1814,8 @@ def _normalize_backmatter_roles_after_boundary(
         page = block.get("page")
         if page is None or page < tail_start or page > tail_boundary.spread_end:
             continue
+        if block.get("role") in _POST_REF_PRESERVE_ROLES:
+            continue
 
         role = block.get("role")
         if backmatter_form == "container":
@@ -2676,6 +2688,22 @@ def _split_merged_heading_blocks(blocks: list[dict], role_profiles: dict | None
     return result
 
 
+def _looks_like_backmatter_body_text(text: str) -> bool:
+    lower = text.lower()
+    markers = (
+        "conflict of interest",
+        "declaration",
+        "publisher",
+        "author contributions",
+        "funding",
+        "acknowledg",
+        "data availability",
+        "supplement",
+        "ethics",
+    )
+    return any(marker in lower for marker in markers)
+
+
 def _exclude_tail_nonref_from_body_flow(blocks: list[dict]) -> None:
     for i, block in enumerate(blocks):
         effective_role = block.get("role")
@@ -2686,18 +2714,13 @@ def _exclude_tail_nonref_from_body_flow(blocks: list[dict]) -> None:
         if block.get("zone") != "tail_nonref_hold_zone":
             continue
 
-        # Veto: body-like paragraph preceded by heading or containing substantial prose
+        # Only convert blocks that carry explicit backmatter evidence;
+        # preserve ordinary body continuation or body-like prose.
         text = str(block.get("text") or block.get("block_content") or "")
-        if i > 0:
-            prev = blocks[i - 1]
-            if str(prev.get("role") or "") in {"section_heading", "subsection_heading"}:
-                continue
-        if len(text) > 120 and ". " in text:
-            lower = text.lower()
-            bp = {"conflict of interest","declaration","publisher","author contributions",
-                  "funding","acknowledg","data availability","supplement"}
-            if not any(b in lower for b in bp):
-                continue
+        if len(text) > 120 and ". " in text and not _looks_like_backmatter_body_text(text):
+            continue
+        if not _looks_like_backmatter_body_text(text):
+            continue
 
         old_role = block.get("role")
         block["role"] = "backmatter_body"
diff --git a/tests/test_ocr_document.py b/tests/test_ocr_document.py
index 4b0e3e0d..19420acb 100644
--- a/tests/test_ocr_document.py
+++ b/tests/test_ocr_document.py
@@ -4371,6 +4371,23 @@ def test_same_page_post_reference_non_reference_block_enters_tail_hold_zone() ->
     assert by_id["ack"]["zone"] == "tail_nonref_hold_zone"
 
 
+def test_tail_nonref_exclusion_does_not_convert_plain_body_prose() -> None:
+    """Plain body prose above same-page References should stay body_paragraph, not backmatter_body."""
+    from paperforge.worker.ocr_document import normalize_document_structure
+
+    blocks = [
+        {"block_id": "h1", "page": 8, "seed_role": "section_heading", "text": "Discussion", "bbox": [80, 120, 320, 180]},
+        {"block_id": "p1", "page": 8, "seed_role": "body_paragraph", "text": "This study demonstrates a reproducible increase in migration after stimulation.", "bbox": [80, 220, 980, 320]},
+        {"block_id": "refs", "page": 8, "seed_role": "reference_heading", "text": "References", "bbox": [80, 1100, 320, 1160]},
+        {"block_id": "r1", "page": 8, "seed_role": "reference_item", "text": "[1] Ref one", "bbox": [80, 1220, 980, 1300]},
+    ]
+
+    _doc, normalized = normalize_document_structure(blocks)
+    by_id = {b["block_id"]: b for b in normalized}
+    assert by_id["p1"]["role"] == "body_paragraph"
+    assert by_id["p1"]["zone"] == "body_zone"
+
+
 # ---------------------------------------------------------------------------
 # Backmatter zone normalization
 # ---------------------------------------------------------------------------