diff --git a/tests/test_ocr_spec_contracts.py b/tests/test_ocr_spec_contracts.py
new file mode 100644
index 00000000..169b30fd
--- /dev/null
+++ b/tests/test_ocr_spec_contracts.py
@@ -0,0 +1,182 @@
+from __future__ import annotations
+
+import pytest
+
+
+class TestOcrRoleContracts:
+    """Architecture contract tests for OCR-v2 role assignment invariants."""
+
+    def test_contract_zone_is_not_role(self) -> None:
+        """Being in the body_zone does not make a block body_paragraph by itself."""
+        from paperforge.worker.ocr_roles import assign_block_role
+
+        block = {
+            "block_label": "text",
+            "block_content": "Check for updates",
+            "block_bbox": [900, 50, 1100, 72],
+        }
+        role = assign_block_role(
+            block, page_blocks=[block], page_width=1200, page_height=1600,
+        )
+        assert role.role not in ("section_heading", "subsection_heading"), (
+            f"zone-only text must not become heading: {role.role}"
+        )
+
+    def test_contract_reference_tail_first(self) -> None:
+        """Reference-like text without reference_zone/numbering support must not become reference_item."""
+        from paperforge.worker.ocr_roles import assign_block_role
+
+        for ordinal_word in (
+            "First,",
+            "Fourth,",
+            "Sixth,",
+            "Additionally,",
+            "Historically,",
+            "Metabolically,",
+        ):
+            block = {
+                "block_label": "text",
+                "block_content": f"{ordinal_word} some body prose that looks like a reference but is not one.",
+                "block_bbox": [100, 200, 800, 260],
+            }
+            role = assign_block_role(
+                block, page_blocks=[block], page_width=1200, page_height=1600,
+            )
+            assert role.role != "reference_item", (
+                f"Ordinal body opening '{ordinal_word}' must not become reference_item, got {role.role}"
+            )
+
+
+class TestOcrRenderContracts:
+    """Architecture contract tests for OCR-v2 renderer invariants."""
+
+    def test_contract_renderer_is_not_semantic_rescue(self) -> None:
+        """render_fulltext_markdown must not promote blocks based on raw text alone."""
+        from paperforge.worker.ocr_render import render_fulltext_markdown
+
+        blocks = [
+            {
+                "block_id": "p1_b1",
+                "page": 1,
+                "role": "body_paragraph",
+                "raw_label": "text",
+                "text": "Fig. 1 This looks like a figure caption but has no accepted figure role.",
+                "bbox": [100, 100, 800, 160],
+                "page_width": 1200,
+                "page_height": 1600,
+            },
+        ]
+        markdown = render_fulltext_markdown(
+            structured_blocks=blocks,
+            resolved_metadata={},
+            figure_inventory={"matched_figures": []},
+            table_inventory={"tables": []},
+            page_count=1,
+            document_structure=None,
+            reader_payload={"reader_figures": [], "consumed_caption_block_ids": []},
+        )
+        assert "Fig. 1" in markdown, "readable text must not be silently dropped"
+        assert "![[assets/figures/" not in markdown, (
+            "renderer must not invent figure embedding from raw text"
+        )
+
+    def test_contract_object_ownership_is_exclusive(self) -> None:
+        """A block consumed by a figure object must not also appear as loose body text."""
+        from paperforge.worker.ocr_render import render_fulltext_markdown
+
+        blocks = [
+            {
+                "block_id": "p1_b1",
+                "page": 1,
+                "role": "figure_caption",
+                "raw_label": "figure_title",
+                "text": "Figure 1. A sample figure.",
+                "bbox": [100, 100, 800, 160],
+                "page_width": 1200,
+                "page_height": 1600,
+            },
+        ]
+        reader_payload = {
+            "reader_figures": [
+                {
+                    "reader_figure_id": "figure_001_reader",
+                    "reader_status": "EXACT_MATCH",
+                    "caption_block_id": "p1_b1",
+                    "consumed_caption_block_ids": ["p1_b1"],
+                },
+            ],
+            "consumed_caption_block_ids": ["p1_b1"],
+        }
+        markdown = render_fulltext_markdown(
+            structured_blocks=blocks,
+            resolved_metadata={},
+            figure_inventory={
+                "matched_figures": [
+                    {
+                        "figure_id": "figure_001",
+                        "text": "Figure 1. A sample figure.",
+                        "page": 1,
+                    },
+                ],
+            },
+            table_inventory={"tables": []},
+            page_count=1,
+            document_structure=None,
+            reader_payload=reader_payload,
+        )
+        occurrences = markdown.count("Figure 1.")
+        assert occurrences == 1, (
+            f"Consumed caption must appear exactly once, found {occurrences} occurrences"
+        )
+
+    def test_contract_reading_segments_are_authoritative(self) -> None:
+        """Conclusion must appear before References in final markdown, regardless of page geometry."""
+        from paperforge.worker.ocr_render import render_fulltext_markdown
+
+        blocks = [
+            {
+                "block_id": "p1_b2",
+                "page": 1,
+                "role": "reference_heading",
+                "raw_label": "paragraph_title",
+                "text": "References",
+                "bbox": [100, 600, 400, 640],
+                "page_width": 1200,
+                "page_height": 1600,
+            },
+            {
+                "block_id": "p1_b3",
+                "page": 1,
+                "role": "reference_item",
+                "raw_label": "text",
+                "text": "[1] Example reference.",
+                "bbox": [100, 660, 800, 700],
+                "page_width": 1200,
+                "page_height": 1600,
+            },
+            {
+                "block_id": "p1_b1",
+                "page": 1,
+                "role": "subsection_heading",
+                "raw_label": "paragraph_title",
+                "text": "Conclusion",
+                "bbox": [100, 800, 400, 840],
+                "page_width": 1200,
+                "page_height": 1600,
+            },
+        ]
+        markdown = render_fulltext_markdown(
+            structured_blocks=blocks,
+            resolved_metadata={},
+            figure_inventory={"matched_figures": []},
+            table_inventory={"tables": []},
+            page_count=1,
+            document_structure=None,
+            reader_payload={"reader_figures": [], "consumed_caption_block_ids": []},
+        )
+        conclusion_pos = markdown.find("Conclusion")
+        references_pos = markdown.find("References")
+        assert conclusion_pos < references_pos, (
+            f"Conclusion must appear before References "
+            f"(conclusion at {conclusion_pos}, refs at {references_pos})"
+        )