diff --git a/src/presa_mano/perception/work_plane.py b/src/presa_mano/perception/work_plane.py
new file mode 100644
index 0000000..af9ca65
--- /dev/null
+++ b/src/presa_mano/perception/work_plane.py
@@ -0,0 +1,146 @@
+"""Work-plane fit, persistence, and below-plane filter.
+
+The work plane is fitted from a user-drawn rectangle on the RGB image. The
+fit uses SVD on the bbox depth cloud (after `nuwa_to_cv_frame`), so the
+plane normal lives in OpenCV RGB-camera frame and is oriented toward the
+camera (negative Z component in OpenCV convention).
+
+The filter `apply_below_plane_filter` removes points more than `offset_mm`
+below the plane: keep `(normal · X + offset) >= -offset_mm/1000`. A
+positive `offset_mm` is permissive (preserves a buffer below the plane).
+"""
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import asdict, dataclass
+from pathlib import Path
+
+import numpy as np
+
+from ..config import data_dir
+
+log = logging.getLogger(__name__)
+
+
+class WorkPlaneError(RuntimeError):
+    """Bad input to `fit_plane_svd` (too few points, degenerate cloud)."""
+
+
+@dataclass
+class WorkPlane:
+    """Plane in CV camera frame. `normal · X + offset = 0`; `normal` is
+    unit-length and oriented toward the camera (negative Z in OpenCV)."""
+    normal: list[float]
+    offset: float
+    centroid: list[float]
+    n_inliers: int
+    fit_rms_mm: float
+    roi: list[int]
+    offset_mm: float
+    created_at: float
+
+
+_WORK_PLANE: WorkPlane | None = None
+_WORK_PLANE_LOADED = False
+
+
+def _work_plane_path() -> Path:
+    return data_dir() / "work_plane.json"
+
+
+def _save_work_plane_to_disk(wp: WorkPlane | None) -> None:
+    path = _work_plane_path()
+    try:
+        if wp is None:
+            if path.exists():
+                path.unlink()
+            return
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text(json.dumps(asdict(wp), indent=2))
+    except Exception as e:
+        log.warning("failed to persist work plane: %s", e)
+
+
+def _load_work_plane_from_disk() -> WorkPlane | None:
+    path = _work_plane_path()
+    try:
+        with open(path, "r") as f:
+            d = json.load(f)
+        d.setdefault("offset_mm", 5.0)   # legacy files
+        return WorkPlane(**d)
+    except FileNotFoundError:
+        return None
+    except Exception as e:
+        log.warning("failed to load persisted work plane (%s)", e)
+        return None
+
+
+def get_work_plane() -> WorkPlane | None:
+    global _WORK_PLANE, _WORK_PLANE_LOADED
+    if not _WORK_PLANE_LOADED:
+        _WORK_PLANE = _load_work_plane_from_disk()
+        _WORK_PLANE_LOADED = True
+    return _WORK_PLANE
+
+
+def set_work_plane(wp: WorkPlane | None) -> None:
+    global _WORK_PLANE, _WORK_PLANE_LOADED
+    _WORK_PLANE = wp
+    _WORK_PLANE_LOADED = True
+    _save_work_plane_to_disk(wp)
+
+
+def fit_plane_svd(points_xyz: np.ndarray
+                  ) -> tuple[np.ndarray, float, np.ndarray, float]:
+    """Least-squares plane via SVD. Returns (normal, offset, centroid, rms_mm).
+
+    Plane equation: `normal · X + offset = 0`. Normal is oriented toward
+    the camera origin (so `signed_distance = normal · X + offset` is
+    positive for points on the camera side of the plane)."""
+    pts = np.asarray(points_xyz, dtype=np.float64)
+    if pts.ndim != 2 or pts.shape[1] < 3:
+        raise WorkPlaneError(f"unexpected points shape for plane fit: {pts.shape}")
+    xyz = pts[:, :3]
+    if len(xyz) < 10:
+        raise WorkPlaneError(f"too few points to fit plane: {len(xyz)}")
+    c = xyz.mean(axis=0)
+    centered = xyz - c
+    _, _, Vt = np.linalg.svd(centered, full_matrices=False)
+    normal = Vt[-1]
+    n = float(np.linalg.norm(normal))
+    if n < 1e-9:
+        raise WorkPlaneError("degenerate plane (zero normal)")
+    normal = normal / n
+    # Orient toward camera origin: signed distance of origin from plane must
+    # be positive, i.e. `-n·c > 0`, i.e. `n·c < 0`.
+    if float(normal @ c) > 0:
+        normal = -normal
+    offset = -float(normal @ c)
+    signed = centered @ normal
+    rms_mm = float(np.sqrt(np.mean(signed * signed)) * 1000.0)
+    return normal, offset, c, rms_mm
+
+
+def apply_below_plane_filter(
+    pts_cv: np.ndarray,
+    wp: WorkPlane | None,
+) -> tuple[np.ndarray, dict]:
+    """Drop points more than `offset_mm` below the plane.
+
+    Returns `(filtered_pts, info)`. `info` always carries:
+        plane_applied (bool): true iff a plane was provided AND there were
+            points to filter
+        n_points_before_plane (int): len(pts_cv) on entry
+    """
+    info = {"plane_applied": False,
+            "n_points_before_plane": int(len(pts_cv))}
+    if wp is None or len(pts_cv) == 0:
+        return pts_cv, info
+    n_vec = np.asarray(wp.normal, dtype=np.float64)
+    d_off = float(wp.offset)
+    offset_m = float(wp.offset_mm) / 1000.0
+    signed = pts_cv @ n_vec + d_off
+    keep = signed >= -offset_m
+    info["plane_applied"] = True
+    return pts_cv[keep], info
diff --git a/tests/test_work_plane.py b/tests/test_work_plane.py
new file mode 100644
index 0000000..af2cd0e
--- /dev/null
+++ b/tests/test_work_plane.py
@@ -0,0 +1,145 @@
+"""Unit tests for perception.work_plane."""
+from __future__ import annotations
+
+import json
+import tempfile
+from pathlib import Path
+from unittest.mock import patch
+
+import numpy as np
+import pytest
+
+from presa_mano.perception.work_plane import (
+    WorkPlane,
+    WorkPlaneError,
+    apply_below_plane_filter,
+    fit_plane_svd,
+    _load_work_plane_from_disk,
+    _save_work_plane_to_disk,
+    _work_plane_path,
+)
+
+
+def _flat_plane_cloud(z: float = 1.0, n: int = 200, noise_m: float = 0.0):
+    """Cloud sampled on the plane z=z (camera frame, OpenCV: camera looks +Z)."""
+    rng = np.random.default_rng(0)
+    xy = rng.uniform(-0.2, 0.2, size=(n, 2))
+    zs = np.full((n, 1), z) + rng.normal(0, noise_m, size=(n, 1))
+    return np.hstack([xy, zs])
+
+
+def test_fit_plane_svd_horizontal_plane():
+    """Plane at z=1.0, normal must point toward camera (negative Z)."""
+    pts = _flat_plane_cloud(z=1.0, n=200, noise_m=0.0)
+    normal, offset, centroid, rms_mm = fit_plane_svd(pts)
+    assert abs(normal[2]) > 0.99
+    assert normal[2] < 0
+    assert abs(offset - 1.0) < 1e-3
+    assert rms_mm < 1.0
+
+
+def test_fit_plane_svd_raises_on_too_few_points():
+    with pytest.raises(WorkPlaneError):
+        fit_plane_svd(np.zeros((5, 3)))
+
+
+def test_apply_below_plane_filter_noop_without_plane():
+    pts = _flat_plane_cloud(z=1.0)
+    out, info = apply_below_plane_filter(pts, None)
+    assert info["plane_applied"] is False
+    assert info["n_points_before_plane"] == len(pts)
+    assert out is pts
+
+
+def test_apply_below_plane_filter_keeps_above_drops_below():
+    """Plane at z=1.0, normal toward camera (-Z): points with z<1.0 are above."""
+    above = _flat_plane_cloud(z=0.95, n=100)
+    below = _flat_plane_cloud(z=1.10, n=100)
+    pts = np.vstack([above, below])
+    wp = WorkPlane(
+        normal=[0.0, 0.0, -1.0], offset=1.0,
+        centroid=[0.0, 0.0, 1.0],
+        n_inliers=100, fit_rms_mm=0.0,
+        roi=[0, 0, 10, 10],
+        offset_mm=0.0,
+        created_at=0.0,
+    )
+    out, info = apply_below_plane_filter(pts, wp)
+    assert info["plane_applied"] is True
+    assert info["n_points_before_plane"] == 200
+    assert len(out) == 100
+    assert np.all(out[:, 2] <= 1.0 + 1e-9)
+
+
+def test_apply_below_plane_filter_offset_buffer_keeps_below():
+    """offset_mm=50 keeps points up to 50mm below the plane."""
+    above = _flat_plane_cloud(z=0.95, n=50)
+    just_below = _flat_plane_cloud(z=1.03, n=50)
+    far_below = _flat_plane_cloud(z=1.20, n=50)
+    pts = np.vstack([above, just_below, far_below])
+    wp = WorkPlane(
+        normal=[0.0, 0.0, -1.0], offset=1.0,
+        centroid=[0.0, 0.0, 1.0],
+        n_inliers=50, fit_rms_mm=0.0,
+        roi=[0, 0, 10, 10],
+        offset_mm=50.0,
+        created_at=0.0,
+    )
+    out, info = apply_below_plane_filter(pts, wp)
+    assert info["plane_applied"] is True
+    assert len(out) == 100   # above (50) + just_below (50); far_below dropped
+
+
+def test_apply_below_plane_filter_empty_input():
+    wp = WorkPlane(
+        normal=[0.0, 0.0, -1.0], offset=1.0,
+        centroid=[0.0, 0.0, 1.0],
+        n_inliers=10, fit_rms_mm=0.0,
+        roi=[0, 0, 10, 10],
+        offset_mm=5.0, created_at=0.0,
+    )
+    out, info = apply_below_plane_filter(np.empty((0, 3)), wp)
+    assert info["plane_applied"] is False   # short-circuit: empty input never "ran"
+    assert info["n_points_before_plane"] == 0
+    assert len(out) == 0
+
+
+def test_work_plane_disk_roundtrip(tmp_path, monkeypatch):
+    """Save then load preserves all fields including offset_mm."""
+    monkeypatch.setenv("PRESA_MANO_DATA", str(tmp_path))
+    wp = WorkPlane(
+        normal=[0.0, 0.0, -1.0], offset=1.0,
+        centroid=[0.1, 0.2, 1.0],
+        n_inliers=42, fit_rms_mm=1.5,
+        roi=[10, 20, 100, 200],
+        offset_mm=7.5,
+        created_at=1234567890.0,
+    )
+    _save_work_plane_to_disk(wp)
+    loaded = _load_work_plane_from_disk()
+    assert loaded is not None
+    assert loaded.normal == wp.normal
+    assert loaded.offset == wp.offset
+    assert loaded.centroid == wp.centroid
+    assert loaded.n_inliers == wp.n_inliers
+    assert loaded.fit_rms_mm == wp.fit_rms_mm
+    assert loaded.roi == wp.roi
+    assert loaded.offset_mm == 7.5
+    assert loaded.created_at == wp.created_at
+
+
+def test_work_plane_load_legacy_file_defaults_offset(tmp_path, monkeypatch):
+    """A pre-existing work_plane.json without offset_mm gets the 5.0 default."""
+    monkeypatch.setenv("PRESA_MANO_DATA", str(tmp_path))
+    path = _work_plane_path()
+    legacy = {
+        "normal": [0.0, 0.0, -1.0], "offset": 1.0,
+        "centroid": [0.0, 0.0, 1.0],
+        "n_inliers": 10, "fit_rms_mm": 1.0,
+        "roi": [0, 0, 10, 10],
+        "created_at": 0.0,
+    }
+    path.write_text(json.dumps(legacy))
+    loaded = _load_work_plane_from_disk()
+    assert loaded is not None
+    assert loaded.offset_mm == 5.0