diff --git a/paperforge/worker/ocr_blocks.py b/paperforge/worker/ocr_blocks.py
index 92b1e8e2..250e14e6 100644
--- a/paperforge/worker/ocr_blocks.py
+++ b/paperforge/worker/ocr_blocks.py
@@ -179,6 +179,37 @@ def build_structured_blocks(
             if resolved.evidence:
                 row.setdefault("evidence", []).extend(resolved.evidence)
 
+    if doc_structure:
+        from paperforge.worker.ocr_document import _apply_zone_labels, infer_zones
+        from paperforge.worker.ocr_families import partition_zone_families
+
+        refreshed_anchors = {
+            "body_family_anchor": getattr(doc_structure, "body_family_anchor", None),
+            "reference_family_anchor": getattr(doc_structure, "reference_family_anchor", None),
+        }
+        refreshed_region_bus = infer_zones(rows, refreshed_anchors)
+        _apply_zone_labels(rows, refreshed_region_bus)
+        partition_zone_families(rows, refreshed_anchors)
+        doc_structure.region_bus = refreshed_region_bus
+
+        refreshed_reference_rows = [row for row in rows if row.get("zone") == "reference_zone"]
+        if refreshed_reference_rows:
+            reference_pages = sorted({int(row.get("page", 0) or 0) for row in refreshed_reference_rows if int(row.get("page", 0) or 0) > 0})
+            if reference_pages:
+                doc_structure.reference_zones = [
+                    {
+                        "page": page,
+                        "column_index": 0,
+                        "y_start": 0,
+                        "y_end": float(next((row.get("page_height", 0) or 0) for row in refreshed_reference_rows if int(row.get("page", 0) or 0) == page) or 0),
+                        "block_indices": [
+                            idx for idx, row in enumerate(rows)
+                            if row.get("zone") == "reference_zone" and int(row.get("page", 0) or 0) == page
+                        ],
+                    }
+                    for page in reference_pages
+                ]
+
     # Sync render_default/index_default after role normalizations
     for row in rows:
         role = row.get("role", "")
diff --git a/paperforge/worker/ocr_document.py b/paperforge/worker/ocr_document.py
index 52662c42..7e5559f1 100644
--- a/paperforge/worker/ocr_document.py
+++ b/paperforge/worker/ocr_document.py
@@ -388,9 +388,9 @@ def _canonical_section_text(block: dict) -> str:
 
 def _is_reference_heading_candidate(block: dict) -> bool:
     marker_type = ((block.get("marker_signature") or {}).get("type") or "none")
-    if marker_type != "canonical_section_name":
+    if _canonical_section_text(block) not in _REFERENCE_ZONE_HEADING_TEXTS:
         return False
-    return _canonical_section_text(block) in _REFERENCE_ZONE_HEADING_TEXTS
+    return marker_type in {"canonical_section_name", "short_fragment", "none"}
 
 
 def _is_reference_item_candidate(block: dict) -> bool:
@@ -3125,6 +3125,32 @@ def normalize_document_structure(blocks: list[dict]) -> tuple[DocumentStructure,
         },
     )
 
+    from paperforge.worker.ocr_roles import resolve_final_role
+
+    family_context = {
+        str(block.get("block_id") or f"block_{idx}"): {
+            "zone": block.get("zone"),
+            "style_family": block.get("style_family"),
+            "style_family_authority": block.get("style_family_authority"),
+        }
+        for idx, block in enumerate(blocks)
+    }
+    anchor_context = {
+        "body_family_anchor": body_family_anchor,
+        "reference_family_anchor": reference_family_anchor,
+    }
+    for block in blocks:
+        resolved = resolve_final_role(block, anchors=anchor_context, families=family_context)
+        if resolved.role != block.get("role"):
+            block["role"] = resolved.role
+            block["role_confidence"] = resolved.confidence
+            if resolved.evidence:
+                block.setdefault("evidence", []).extend(resolved.evidence)
+
+    region_bus = infer_zones(blocks, anchor_context)
+    _apply_zone_labels(blocks, region_bus)
+    partition_zone_families(blocks, anchor_context)
+
     tail_spread = _reconcile_tail_spread(blocks, page_layouts)
     if tail_spread is not None:
         backmatter_form = _classify_backmatter_form(tail_spread, blocks)
diff --git a/paperforge/worker/ocr_families.py b/paperforge/worker/ocr_families.py
index 6de3c109..7efe8c46 100644
--- a/paperforge/worker/ocr_families.py
+++ b/paperforge/worker/ocr_families.py
@@ -509,7 +509,7 @@ def _has_reference_heading_near_family(blocks: list[dict], family_pages: list[in
             continue
         marker_type = ((block.get("marker_signature") or {}).get("type") or "none")
         text = str(block.get("text") or block.get("block_content") or "").strip().lower()
-        if marker_type == "canonical_section_name" and text in _REFERENCE_HEADING_TEXTS:
+        if marker_type in {"canonical_section_name", "short_fragment", "none"} and text in _REFERENCE_HEADING_TEXTS:
             return True
     return False
 
@@ -527,7 +527,8 @@ def _build_reference_family_candidate(
     font_family, font_size_bucket, width_bucket, x_center_bucket = family_key
     strong_tail = tail_continuity >= 1.0
     strong_marker_family = item_count >= 2 and marker_count >= 2
-    can_accept = strong_marker_family and strong_tail
+    heading_supported_singleton = heading_present and item_count >= 1 and marker_count >= 1
+    can_accept = strong_tail and (strong_marker_family or heading_supported_singleton)
     reason = "reference_markers_with_tail_continuity"
     if heading_present:
         reason = f"{reason}_and_heading_binding"
diff --git a/paperforge/worker/ocr_figures.py b/paperforge/worker/ocr_figures.py
index 10c5de2e..133cc278 100644
--- a/paperforge/worker/ocr_figures.py
+++ b/paperforge/worker/ocr_figures.py
@@ -122,7 +122,11 @@ def _looks_like_figure_narrative_prose(text: str) -> bool:
 
 
 def _is_body_mention(block: dict) -> bool:
+    zone = str(block.get("zone") or "")
+    style_family = str(block.get("style_family") or "")
     raw_role = block.get("raw_role", block.get("role", ""))
+    if zone == "display_zone" and style_family == "legend_like":
+        return False
     if raw_role == "body_paragraph":
         return True
     if raw_role == "figure_caption_candidate":
diff --git a/paperforge/worker/ocr_roles.py b/paperforge/worker/ocr_roles.py
index 3d8b8b76..74572e12 100644
--- a/paperforge/worker/ocr_roles.py
+++ b/paperforge/worker/ocr_roles.py
@@ -434,11 +434,34 @@ def resolve_final_role(
     current_role = str(block.get("role") or "body_paragraph")
     current_confidence = float(block.get("role_confidence") or 0.6)
     body_anchor = anchors.get("body_family_anchor") or {}
+    reference_anchor = anchors.get("reference_family_anchor") or {}
     body_anchor_accepted = str(body_anchor.get("status") or "").upper() == "ACCEPT"
+    reference_anchor_accepted = str(reference_anchor.get("status") or "").upper() == "ACCEPT"
     in_body_zone = zone == "body_zone"
     strong_legend_authority = style_family_authority in {"figure_marker", "figure_family_anchor"}
 
     if current_role == "body_paragraph":
+        if (
+            zone == "reference_zone"
+            and reference_anchor_accepted
+            and style_family == "reference_like"
+            and marker_type in {
+                "reference_numeric_bracket",
+                "reference_numeric_dot",
+                "reference_numeric_parenthesis",
+                "reference_pattern",
+                "citation_line",
+            }
+        ):
+            return RoleAssignment(
+                role="reference_item",
+                confidence=max(current_confidence, 0.82),
+                evidence=[
+                    "late role resolution: reference_like family + reference zone",
+                    f"style_family_authority={style_family_authority or 'none'}",
+                    f"context_source={context_source}",
+                ],
+            )
         if (
             in_body_zone
             and body_anchor_accepted
@@ -458,6 +481,22 @@ def resolve_final_role(
                     f"context_source={context_source}",
                 ],
             )
+        if (
+            zone == "display_zone"
+            and strong_legend_authority
+            and style_family == "legend_like"
+            and marker_type == "figure_number"
+            and not _looks_like_late_figure_narrative_prose(str(block.get("text") or ""))
+        ):
+            return RoleAssignment(
+                role="figure_caption_candidate",
+                confidence=max(current_confidence, 0.8),
+                evidence=[
+                    "late role resolution: display-zone legend candidate from figure family",
+                    f"style_family_authority={style_family_authority or 'none'}",
+                    f"context_source={context_source}",
+                ],
+            )
 
     return RoleAssignment(
         role=current_role,
diff --git a/tests/test_ocr_document.py b/tests/test_ocr_document.py
index e19d3335..d978d76f 100644
--- a/tests/test_ocr_document.py
+++ b/tests/test_ocr_document.py
@@ -305,6 +305,56 @@ def test_normalize_document_structure_wires_style_family_artifacts_into_blocks()
     assert normalized_blocks[3]["style_family"] == "legend_like"
 
 
+def test_normalize_document_structure_preserves_reference_zone_integrity_from_anchor_flow() -> None:
+    from paperforge.worker.ocr_document import normalize_document_structure
+
+    blocks = [
+        {
+            "block_id": "p2_b1",
+            "page": 2,
+            "role": "body_paragraph",
+            "text": "Stable body paragraph with repeated typography and enough prose to form the body family anchor. " * 2,
+            "bbox": [100, 120, 460, 280],
+            "page_width": 1200,
+            "page_height": 1600,
+            "marker_signature": {"type": "none"},
+            "span_signature": {"font_size_median": 9.0, "font_size_bucket": 9.0, "font_family_norm": "Times"},
+            "layout_signature": {"width": 260, "width_bucket": 250, "x_center": 230, "x_center_bucket": 250},
+        },
+        {
+            "block_id": "p5_b1",
+            "page": 5,
+            "role": "reference_heading",
+            "text": "References",
+            "bbox": [100, 100, 340, 140],
+            "page_width": 1200,
+            "page_height": 1600,
+            "marker_signature": {"type": "canonical_section_name"},
+            "span_signature": {"font_size_median": 10.0, "font_family_norm": "Times"},
+            "layout_signature": {"width": 120, "x_center": 220},
+        },
+        {
+            "block_id": "p5_b2",
+            "page": 5,
+            "role": "body_paragraph",
+            "text": "[1] Example reference with authors and journal details.",
+            "bbox": [110, 180, 500, 260],
+            "page_width": 1200,
+            "page_height": 1600,
+            "marker_signature": {"type": "reference_numeric_bracket", "number": 1},
+            "span_signature": {"font_size_median": 8.5, "font_size_bucket": 8.5, "font_family_norm": "Times"},
+            "layout_signature": {"width": 250, "width_bucket": 250, "x_center": 240, "x_center_bucket": 250},
+        },
+    ]
+
+    doc, normalized_blocks = normalize_document_structure(blocks)
+
+    assert doc.reference_zones
+    ref_row = next(block for block in normalized_blocks if block["block_id"] == "p5_b2")
+    assert ref_row.get("zone") == "reference_zone"
+    assert ref_row.get("style_family") == "reference_like"
+
+
 def test_candidate_resolution_demotes_body_spine_narrative_figure_mentions() -> None:
     from paperforge.worker.ocr_document import DocumentStructure, PageLayoutProfile, _resolve_ambiguous_candidates
 
diff --git a/tests/test_ocr_figures.py b/tests/test_ocr_figures.py
index f10c54c5..b4299baf 100644
--- a/tests/test_ocr_figures.py
+++ b/tests/test_ocr_figures.py
@@ -228,6 +228,56 @@ def test_formal_figure_count_is_based_on_legends_not_raw_images() -> None:
     assert len(inventory["figure_legends"]) == 1
 
 
+def test_validation_first_legend_does_not_promote_body_figure_mention() -> None:
+    from paperforge.worker.ocr_figures import build_figure_inventory
+
+    structured_blocks = [
+        {
+            "paper_id": "KEY010B",
+            "page": 3,
+            "block_id": "p3_b1",
+            "role": "body_paragraph",
+            "raw_role": "body_paragraph",
+            "raw_label": "text",
+            "text": "Figure 2 shows the comparative response over time and this sentence continues as narrative body prose for the main discussion section.",
+            "bbox": [90, 100, 530, 180],
+            "zone": "body_zone",
+            "style_family": "body_like",
+            "style_family_authority": "body_family_anchor",
+            "body_spine_member": True,
+            "marker_signature": {"type": "figure_number", "number": 2},
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 3,
+            "block_id": "p3_b2",
+            "role": "figure_asset",
+            "raw_label": "image",
+            "text": "",
+            "bbox": [620, 140, 1030, 520],
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 3,
+            "block_id": "p3_b3",
+            "role": "body_paragraph",
+            "raw_label": "text",
+            "text": "Figure 2. Formal caption with sufficient descriptive text to support validation-first legend matching near the media asset.",
+            "bbox": [620, 540, 1040, 620],
+            "zone": "display_zone",
+            "style_family": "legend_like",
+            "style_family_authority": "figure_family_anchor",
+            "marker_signature": {"type": "figure_number", "number": 2},
+        },
+    ]
+
+    inventory = build_figure_inventory(structured_blocks)
+
+    assert all(legend.get("legend_block_id", legend.get("block_id")) != "p3_b1" for legend in inventory["figure_legends"])
+    assert any(legend.get("legend_block_id", legend.get("block_id")) == "p3_b3" for legend in inventory["figure_legends"])
+    assert inventory["official_figure_count"] == 1
+
+
 def test_figure_inventory_includes_all_sections() -> None:
     from paperforge.worker.ocr_figures import build_figure_inventory
 
diff --git a/tests/test_ocr_health.py b/tests/test_ocr_health.py
index f70acd6c..a6262fee 100644
--- a/tests/test_ocr_health.py
+++ b/tests/test_ocr_health.py
@@ -154,6 +154,37 @@ def test_ocr_health_includes_decision_counts() -> None:
     assert report["role_rescue_count"] == 1
 
 
+def test_ocr_health_reports_anchor_first_authority_summaries() -> None:
+    from paperforge.worker.ocr_document import DocumentStructure
+    from paperforge.worker.ocr_health import build_ocr_health
+
+    doc = DocumentStructure(
+        body_family_anchor={"status": "ACCEPT"},
+        reference_family_anchor={"status": "ACCEPT"},
+        region_bus={
+            "body_zone": {"status": "ACCEPT", "block_ids": ["p2_b1"]},
+            "reference_zone": {"status": "ACCEPT", "block_ids": ["p5_b2"]},
+        },
+    )
+    blocks = [
+        {"block_id": "p2_b1", "role": "body_paragraph", "text": "Body", "page": 2, "bbox": [0, 0, 10, 10]},
+        {"block_id": "p5_b2", "role": "reference_item", "text": "[1] Ref", "page": 5, "bbox": [0, 0, 10, 10]},
+    ]
+
+    report = build_ocr_health(
+        page_count=5,
+        raw_blocks_count=2,
+        structured_blocks=blocks,
+        figure_inventory={},
+        table_inventory={},
+        doc_structure=doc,
+    )
+
+    assert report["anchor_summary"]["body_family_anchor"] == "ACCEPT"
+    assert report["anchor_summary"]["reference_family_anchor"] == "ACCEPT"
+    assert report["zone_summary"]["reference_zone"] == "ACCEPT"
+
+
 def test_ocr_health_includes_tail_boundary_confidence() -> None:
     from paperforge.worker.ocr_health import build_ocr_health
 
diff --git a/tests/test_ocr_real_paper_regressions.py b/tests/test_ocr_real_paper_regressions.py
index da2e2975..b7fbd700 100644
--- a/tests/test_ocr_real_paper_regressions.py
+++ b/tests/test_ocr_real_paper_regressions.py
@@ -185,3 +185,19 @@ def test_control_papers_keep_body_and_tail_stability(rebuilt_real_papers: dict,
 
     assert roles.count("body_paragraph") >= CONTROL_MIN_BODY[key]
     assert any(w in fulltext for w in ["References", "references", "REFERENCE", "Bibliography"]), "Tail reference marker not found"
+
+
+@pytest.mark.parametrize("key", PROBLEM_KEYS)
+def test_problem_papers_keep_reference_roles_and_exclude_legend_family_from_body(
+    rebuilt_real_papers: dict,
+    _ocr_root: Path,
+    key: str,
+) -> None:
+    _require_artifacts(_ocr_root, key)
+    blocks = _read_jsonl(_structured_path(_ocr_root, key))
+
+    assert any(block.get("role") == "reference_item" for block in blocks)
+    assert not any(
+        block.get("role") == "body_paragraph" and block.get("style_family") in {"legend_like", "table_caption_like", "reference_like"}
+        for block in blocks
+    )
diff --git a/tests/test_ocr_roles.py b/tests/test_ocr_roles.py
index 46077aa9..30fe8990 100644
--- a/tests/test_ocr_roles.py
+++ b/tests/test_ocr_roles.py
@@ -256,6 +256,92 @@ def test_body_with_noise_phrase_below_backmatter_heading() -> None:
     )
 
 
+def test_pipeline_keeps_reference_zone_and_legend_family_out_of_default_body() -> None:
+    from paperforge.worker.ocr_blocks import build_structured_blocks
+
+    raw_blocks = [
+        {
+            "paper_id": "KEY010B",
+            "page": 2,
+            "block_id": "p2_b1",
+            "raw_label": "text",
+            "raw_order": 0,
+            "bbox": [110, 120, 470, 280],
+            "text": "Main body paragraph with stable narrative prose and enough repeated content to establish the body family anchor for the document. " * 2,
+            "page_width": 1200,
+            "page_height": 1600,
+            "span_metadata": {"size": 9.0, "font": "Times", "flags": ""},
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 3,
+            "block_id": "p3_b1",
+            "raw_label": "text",
+            "raw_order": 0,
+            "bbox": [112, 118, 472, 282],
+            "text": "Another core body paragraph repeating the same typography and width so the anchor-first body family remains dominant across the middle pages. " * 2,
+            "page_width": 1200,
+            "page_height": 1600,
+            "span_metadata": {"size": 9.0, "font": "Times", "flags": ""},
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 5,
+            "block_id": "p5_b1",
+            "raw_label": "paragraph_title",
+            "raw_order": 0,
+            "bbox": [100, 80, 350, 120],
+            "text": "References",
+            "page_width": 1200,
+            "page_height": 1600,
+            "span_metadata": {"size": 11.0, "font": "Times", "flags": "bold"},
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 5,
+            "block_id": "p5_b2",
+            "raw_label": "text",
+            "raw_order": 1,
+            "bbox": [110, 170, 500, 260],
+            "text": "[1] Example reference entry with journal and year details.",
+            "page_width": 1200,
+            "page_height": 1600,
+            "span_metadata": {"size": 8.5, "font": "Times", "flags": ""},
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 4,
+            "block_id": "p4_b9",
+            "raw_label": "text",
+            "raw_order": 2,
+            "bbox": [720, 780, 1030, 860],
+            "text": "Figure 1. Compact legend text for the nearby display panel.",
+            "page_width": 1200,
+            "page_height": 1600,
+            "span_metadata": {"size": 8.0, "font": "Times", "flags": ""},
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 4,
+            "block_id": "p4_b10",
+            "raw_label": "figure",
+            "raw_order": 3,
+            "bbox": [700, 420, 1040, 760],
+            "text": "",
+            "page_width": 1200,
+            "page_height": 1600,
+        },
+    ]
+
+    rows, _ = build_structured_blocks(raw_blocks)
+
+    assert any(row["role"] == "reference_item" for row in rows)
+    assert not any(
+        row["role"] == "body_paragraph" and row.get("style_family") == "legend_like"
+        for row in rows
+    )
+
+
 def test_style_aware_unnumbered_heading_detection() -> None:
     """Unnumbered headings with distinct visual style are detected as headings.
 
diff --git a/tests/test_ocr_tables.py b/tests/test_ocr_tables.py
index 561eae28..f878ac66 100644
--- a/tests/test_ocr_tables.py
+++ b/tests/test_ocr_tables.py
@@ -62,6 +62,40 @@ def test_table_without_asset_is_tracked_as_unmatched_caption() -> None:
     assert len(inventory["unmatched_captions"]) == 1
 
 
+def test_validation_first_table_candidate_remains_stable_under_anchor_first_roles() -> None:
+    from paperforge.worker.ocr_tables import build_table_inventory
+
+    structured_blocks = [
+        {
+            "paper_id": "KEY010B",
+            "page": 6,
+            "block_id": "p6_a1",
+            "role": "table_asset",
+            "raw_label": "table",
+            "text": "parsed table image",
+            "bbox": [120, 140, 760, 520],
+        },
+        {
+            "paper_id": "KEY010B",
+            "page": 6,
+            "block_id": "p6_c1",
+            "role": "body_paragraph",
+            "raw_label": "text",
+            "text": "Table 4. Quantitative outcomes for anchor-first validation.",
+            "bbox": [120, 540, 760, 600],
+            "zone": "display_zone",
+            "style_family": "table_caption_like",
+            "marker_signature": {"type": "table_number", "number": 4},
+        },
+    ]
+
+    inventory = build_table_inventory(structured_blocks)
+
+    assert inventory["official_table_count"] == 1
+    assert inventory["tables"][0]["caption_block_id"] == "p6_c1"
+    assert inventory["tables"][0]["has_asset"] is True
+
+
 def test_continuation_table_matches_same_page_asset() -> None:
     from paperforge.worker.ocr_tables import build_table_inventory