diff --git a/tests/test_perception_provider.py b/tests/test_perception_provider.py
new file mode 100644
index 0000000..6f93208
--- /dev/null
+++ b/tests/test_perception_provider.py
@@ -0,0 +1,79 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+from core.models import Scene
+from perception.ocr import OCRBox
+from perception.provider import DefaultPerceptionProvider, NullPerceptionProvider
+from perception.scene_builder import build_scene as build_fused_scene
+from perception.scene_builder import infer_png_size
+from perception.ui_parser import parse_ui_tree
+
+PNG_10X20 = (
+    b"\x89PNG\r\n\x1a\n"
+    b"\x00\x00\x00\r"
+    b"IHDR"
+    b"\x00\x00\x00\x0a"
+    b"\x00\x00\x00\x14"
+)
+
+TREE_XML = """
+<AppiumAUT x="0" y="0" width="10" height="20">
+  <XCUIElementTypeButton name="Search" label="Search" x="1" y="2" width="4" height="4" />
+  <XCUIElementTypeImage name="Settings" label="Settings" x="6" y="2" width="3" height="3" />
+</AppiumAUT>
+"""
+
+
+class FakeOCREngine:
+    """Stand-in for PaddleOCREngine that never touches the real dependency."""
+
+    def __init__(self, boxes: list[OCRBox]) -> None:
+        self._boxes = boxes
+        self.calls: list[object] = []
+
+    def extract(self, image: bytes | str | Path) -> list[OCRBox]:
+        self.calls.append(image)
+        return self._boxes
+
+
+def test_null_perception_provider_returns_empty_scene_without_ocr() -> None:
+    provider = NullPerceptionProvider()
+
+    scene = provider.build_scene(PNG_10X20, TREE_XML)
+
+    width, height = infer_png_size(PNG_10X20)
+    assert scene == Scene(width=width, height=height, elements=[])
+    assert width == 10
+    assert height == 20
+    assert scene.elements == []
+
+
+def test_null_perception_provider_infers_zero_size_for_non_png() -> None:
+    provider = NullPerceptionProvider()
+
+    scene = provider.build_scene(b"not-a-png", None)
+
+    assert scene == Scene(width=0, height=0, elements=[])
+
+
+def test_default_perception_provider_wires_through_scene_builder() -> None:
+    fake_boxes = [
+        OCRBox(text="Search", bounds=parse_ui_tree(TREE_XML)[0].bounds, confidence=0.9),
+    ]
+    fake_engine = FakeOCREngine(fake_boxes)
+    provider = DefaultPerceptionProvider(ocr_engine=fake_engine)  # type: ignore[arg-type]
+
+    scene = provider.build_scene(PNG_10X20, TREE_XML)
+
+    width, height = infer_png_size(PNG_10X20)
+    expected = build_fused_scene(
+        screen_width=width,
+        screen_height=height,
+        ui_elements=parse_ui_tree(TREE_XML),
+        ocr_elements=[fake_boxes[0].to_scene_element("ocr-000")],
+    )
+
+    assert scene == expected
+    assert len(scene.elements) == len(parse_ui_tree(TREE_XML))
+    assert fake_engine.calls == [PNG_10X20]