diff --git a/README.md b/README.md
index 3b73001..4585520 100644
--- a/README.md
+++ b/README.md
@@ -96,12 +96,12 @@ Speckle Model
 | File | Purpose |
 |------|---------|
 | `main.py` | Entry point, orchestrates the full pipeline |
-| `utils/helpers.py` | Shared utilities: safe attribute access (`_get`) and unit scale constants |
+| `utils/helpers.py` | Shared utilities: safe attribute access (`_get`), unit scale constants, and `resolve_scale` |
 | `utils/traversal.py` | Walks the Speckle collection tree (Root > Collection* > DataObject) |
 | `utils/mapper.py` | Reads IFC entity class from `properties.Attributes.type` |
 | `utils/geometry.py` | Converts Speckle Mesh/Brep/BrepX geometry to IfcPolygonalFaceSet |
 | `utils/curves.py` | Converts Speckle 2D curve geometry (Polycurve, Line, Arc) to IfcIndexedPolyCurve |
-| `utils/instances.py` | Handles InstanceProxy objects with shared geometry (IfcMappedItem) |
+| `utils/instances.py` | Handles InstanceProxy objects with shared geometry (IfcMappedItem), content-based geometry dedup |
 | `utils/properties.py` | Clones all properties, quantities, and attributes into IFC entities |
 | `utils/type_manager.py` | Creates and caches IfcTypeObjects, supports both explicit and derived type classes |
 | `utils/materials.py` | Maps Speckle render materials to IfcSurfaceStyle colours |
@@ -150,21 +150,25 @@ Curves are typically found wrapped inside `DataObject.displayValue`, following t
 1. Extract vertices and faces from each mesh in `displayValue`
 2. Scale vertices to millimetres based on the mesh's unit declaration
 3. Deduplicate vertices via snap grid (0.01mm tolerance) to avoid IFC GEM111 errors
-4. Build `IfcPolygonalFaceSet` with `IfcCartesianPointList3D` + `IfcIndexedPolygonalFace`
-5. Compute bounding box origin for `IfcLocalPlacement`, offset vertices relative to it
+4. Round coordinates to 0.001mm precision for compact IFC file output
+5. Build `IfcPolygonalFaceSet` with `IfcCartesianPointList3D` + `IfcIndexedPolygonalFace`
+6. Compute bounding box origin for `IfcLocalPlacement`, offset vertices relative to it
 
 ### 2D Curve Conversion
 
 1. Extract curve segments from the object or its `displayValue`
 2. Parse each segment type (Line → start/end, Arc → start/mid/end, Polyline → point sequence)
 3. Deduplicate points via snap grid (0.01mm tolerance)
-4. Build `IfcIndexedPolyCurve` with `IfcCartesianPointList3D` + `IfcLineIndex` / `IfcArcIndex` segments
-5. Compute bounding box origin for placement, offset points relative to it
+4. Round coordinates to 0.001mm precision for compact IFC file output
+5. Build `IfcIndexedPolyCurve` with `IfcCartesianPointList3D` + `IfcLineIndex` / `IfcArcIndex` segments
+6. Compute bounding box origin for placement, offset points relative to it
 
 ### Instance Objects (Path A / B2)
 
 Speckle `InstanceProxy` objects reference shared definition geometry via `definitionId`. Geometry is built once as an `IfcRepresentationMap`, then each instance references it via `IfcMappedItem` + `IfcCartesianTransformationOperator3DnonUniform`. This avoids duplicating vertex/curve data across hundreds of identical elements. Both mesh and curve definitions are supported.
 
+**Content-based geometry deduplication**: Instance definitions with identical vertex/face data and materials are detected via MD5 content hashing and share a single `IfcRepresentationMap`, even if they have different `definitionId`s. Direction vectors for transform operators are also cached and reused across instances.
+
 ## Material Handling
 
 Materials are read from `root.renderMaterialProxies` and applied as `IfcSurfaceStyle` on geometry items. Each proxy contains a `RenderMaterial` (name, diffuse colour as ARGB packed int, opacity) and a list of object references.
diff --git a/main.py b/main.py
index ed02ba5..728fd22 100644
--- a/main.py
+++ b/main.py
@@ -1,3 +1,4 @@
+import zipfile
 from datetime import datetime
 
 import ifcopenshell.api
@@ -217,11 +218,17 @@ def automate_function(
 
     ifc.write(ifc_filename)
     print(f"\nIFC file written: {ifc_filename}")
+    
+    zip_filename = f"{file_name}_{timestamp}.zip"
+    with zipfile.ZipFile(zip_filename, "w", zipfile.ZIP_DEFLATED) as zf:
+        zf.write(ifc_filename)
+    print(f"Zipped: {zip_filename}")
+
     try:
-        automate_context.mark_run_success("Success! You can download the IF file below.")
-        automate_context.store_file_result(f"./{ifc_filename}")
+        automate_context.mark_run_success("Success! You can download the IFC file below.")
+        automate_context.store_file_result(f"./{zip_filename}")
     except Exception as e:
-        print(f"  ⚠️  Could not upload file result (network issue?): {e}")
+        print(f"  Could not upload file result (network issue?): {e}")
         automate_context.mark_run_failed(f"Something went wrong when storing file result. Exception detail: {e}") 
 
     print(f"\n{'=' * 60}")
diff --git a/sample_files/geometry.py b/sample_files/geometry.py
new file mode 100644
index 0000000..cfc1404
--- /dev/null
+++ b/sample_files/geometry.py
@@ -0,0 +1,447 @@
+# =============================================================================
+# geometry.py
+# Converts Speckle DataObject geometry → IFC IfcPolygonalFaceSet + IfcLocalPlacement
+#
+# Key facts:
+#   - After specklepy receive(), vertices and faces are FLAT Python lists
+#   - displayValue is an array of Mesh objects
+#   - Units are in mm (for Revit), scale to metres for IFC
+#   - Vertices are in absolute world coordinates
+#   - Uses IfcPolygonalFaceSet (indexed vertices) instead of IfcFacetedBrep
+#     for compact output — each vertex stored once, not once per face.
+# =============================================================================
+
+import ifcopenshell
+from specklepy.objects.base import Base
+from utils.helpers import _get, MM_SCALES as _UNIT_SCALES
+
+
+# --------------------------------------------------------------------------- #
+# Geometry validation helpers (GEM111 fix)
+# --------------------------------------------------------------------------- #
+
+# Minimum distance in mm below which two vertices are considered identical (GEM111).
+_VERTEX_MERGE_TOL = 0.01  # 0.01 mm
+_INV_TOL = 1.0 / _VERTEX_MERGE_TOL  # pre-computed: multiply instead of divide
+
+
+def build_ifc_facesets(ifc, verts_scaled: list, face_groups: list) -> list:
+    """
+    Build a list of IfcPolygonalFaceSet from scaled (x,y,z) vertices and face index groups.
+
+    Uses IfcCartesianPointList3D + IfcIndexedPolygonalFace for compact output.
+    Vertices are deduplicated via snap grid so each unique position is stored once.
+
+    GEM111 fix: skip faces with near-duplicate vertices (snapped to same grid cell).
+
+    verts_scaled: flat list of already-scaled floats [x0,y0,z0, x1,y1,z1, ...]
+    face_groups:  list of index lists [[i,j,k], [i,j,k,l], ...]
+    Returns: list of IfcPolygonalFaceSet (typically one, empty on failure).
+    """
+    snap_to_idx = {}   # snap_key → 0-based index in deduped_verts
+    deduped_verts = [] # [[x, y, z], ...] — lists for direct IFC use
+    inv_tol = _INV_TOL
+
+    # Validate faces and remap indices to deduplicated vertex list
+    valid_faces = []  # list of (idx0+1, idx1+1, ...) tuples (1-based for IFC)
+    for indices in face_groups:
+        try:
+            remapped = []
+            seen_snaps = set()
+            degenerate = False
+
+            for i in indices:
+                i3 = i * 3
+                x = verts_scaled[i3]
+                y = verts_scaled[i3 + 1]
+                z = verts_scaled[i3 + 2]
+                key = (round(x * inv_tol), round(y * inv_tol), round(z * inv_tol))
+                if key in seen_snaps:
+                    degenerate = True
+                    break
+                seen_snaps.add(key)
+                idx = snap_to_idx.get(key)
+                if idx is None:
+                    idx = len(deduped_verts)
+                    snap_to_idx[key] = idx
+                    deduped_verts.append([x, y, z])
+                remapped.append(idx + 1)  # 1-based for IFC
+
+            if degenerate or len(remapped) < 3:
+                continue
+            valid_faces.append(remapped)
+        except Exception:
+            continue
+
+    if not valid_faces or not deduped_verts:
+        return []
+
+    # Round vertex coordinates to reduce IFC text file size
+    # 3 decimal places = 0.001mm precision (more than sufficient)
+    for v in deduped_verts:
+        v[0] = round(v[0], 3)
+        v[1] = round(v[1], 3)
+        v[2] = round(v[2], 3)
+
+    # Build IFC entities
+    try:
+        point_list = ifc.createIfcCartesianPointList3D(deduped_verts)
+        ifc_faces = [
+            ifc.createIfcIndexedPolygonalFace(fi) for fi in valid_faces
+        ]
+        faceset = ifc.createIfcPolygonalFaceSet(point_list, None, ifc_faces, None)
+        return [faceset]
+    except Exception:
+        return []
+
+
+def unwrap_chunks(raw) -> list:
+    """
+    Flatten a Speckle data array into a plain Python list of numbers.
+
+    Handles two cases:
+      1. Already flat list of numbers (after specklepy receive deserializes)
+         → returned as-is (fast path)
+      2. List of DataChunk objects (raw from server before deserialization)
+         → each chunk's .data list is concatenated
+    """
+    if not raw:
+        return []
+
+    # Fast path: if first item is a number, assume all items are numbers
+    first = raw[0]
+    if isinstance(first, (int, float)):
+        return raw
+
+    # Slow path: DataChunk objects or mixed content
+    result = []
+    for item in raw:
+        if item is None:
+            continue
+        if isinstance(item, (int, float)):
+            result.append(item)
+            continue
+        speckle_type = getattr(item, "speckle_type", "") or ""
+        if "DataChunk" in speckle_type:
+            chunk_data = _get(item, "data") or _get(item, "@data")
+            if chunk_data:
+                result.extend(list(chunk_data))
+        else:
+            try:
+                result.extend(list(item))
+            except Exception:
+                pass
+    return result
+
+
+def _resolve_scale(obj, stream_scale: float) -> float:
+    """Resolve unit scale: obj.units → stream fallback."""
+    units = _get(obj, "units")
+    if units and isinstance(units, str):
+        return _UNIT_SCALES.get(units.lower().strip(), stream_scale)
+    return stream_scale
+
+
+# --------------------------------------------------------------------------- #
+# Mesh extraction
+# --------------------------------------------------------------------------- #
+
+def _is_mesh(item) -> bool:
+    """
+    Detect if a specklepy object is a Mesh.
+    Uses speckle_type string — more reliable than hasattr on Base objects.
+    """
+    if item is None:
+        return False
+    speckle_type = _get(item, "speckle_type") or ""
+    if "Mesh" in speckle_type:
+        return True
+    # Fallback: has both vertices and faces data
+    verts = _get(item, "vertices")
+    faces = _get(item, "faces")
+    return verts is not None and faces is not None
+
+
+def _collect_meshes_from_display(obj) -> list:
+    """
+    Collect Mesh objects from an object's displayValue.
+    If an item is not a Mesh (e.g. BrepX, Brep), recursively check
+    its own displayValue for nested meshes.
+    """
+    meshes = []
+    for key in ["displayValue", "@displayValue", "_displayValue"]:
+        display = _get(obj, key)
+        if display is None:
+            continue
+        items = display if isinstance(display, list) else [display]
+        for item in items:
+            if item is None:
+                continue
+            if _is_mesh(item):
+                meshes.append(item)
+            else:
+                # BrepX / Brep / other geometry types may carry a nested
+                # displayValue with the tessellated mesh representation
+                meshes.extend(_collect_meshes_from_display(item))
+        if meshes:
+            break
+    return meshes
+
+
+def get_display_meshes(obj: Base) -> list:
+    """
+    Extract all Mesh objects from a DataObject's displayValue.
+    Handles nested geometry types (BrepX, Brep) that wrap meshes
+    inside their own displayValue.
+    """
+    meshes = _collect_meshes_from_display(obj)
+
+    # Fallback: object itself is a Mesh
+    if not meshes and _is_mesh(obj):
+        speckle_type = _get(obj, "speckle_type") or ""
+        if "Mesh" in speckle_type:
+            meshes.append(obj)
+
+    return meshes
+
+
+def get_display_instances(obj: Base) -> list:
+    """
+    Extract InstanceProxy objects from a DataObject's displayValue.
+
+    Per the official speckleifc converter, every IFC element's displayValue
+    contains InstanceProxy objects (not raw meshes). Each InstanceProxy has:
+      - transform:    16-float row-major matrix, translation in metres
+      - definitionId: "DEFINITION:{meshAppId}" string
+      - units:        "m"
+
+    Raw meshes do NOT appear in displayValue in IFC→Speckle exports.
+    """
+    instances = []
+    for key in ["displayValue", "@displayValue", "_displayValue"]:
+        display = _get(obj, key)
+        if display is None:
+            continue
+        items = display if isinstance(display, list) else [display]
+        for item in items:
+            if item is None:
+                continue
+            transform     = _get(item, "transform")
+            definition_id = _get(item, "definitionId")
+            if transform is not None and definition_id is not None:
+                instances.append(item)
+        if instances:
+            break
+    return instances
+
+
+# --------------------------------------------------------------------------- #
+# Face decoding
+# --------------------------------------------------------------------------- #
+
+def decode_faces(faces_raw: list) -> list:
+    """
+    Decode Speckle's run-length encoded face list into vertex index groups.
+    Format: [n, i0, i1, ..., n, i0, i1, ...]
+      n=0 → triangle (legacy), n=1 → quad (legacy), n≥3 → n-gon
+    """
+    decoded = []
+    i = 0
+    total = len(faces_raw)
+    # Check if values are already ints (common after unwrap_chunks)
+    already_int = total > 0 and isinstance(faces_raw[0], int)
+    while i < total:
+        n = faces_raw[i] if already_int else int(faces_raw[i])
+        if n == 0:
+            n = 3
+        elif n == 1:
+            n = 4
+        end = i + 1 + n
+        if end > total:
+            break
+        if already_int:
+            decoded.append(faces_raw[i + 1:end])
+        else:
+            decoded.append([int(v) for v in faces_raw[i + 1:end]])
+        i = end
+    return decoded
+
+
+# --------------------------------------------------------------------------- #
+# Bounding box + placement
+# --------------------------------------------------------------------------- #
+
+def compute_origin(flat_verts: list) -> tuple:
+    """
+    Compute placement origin from scaled vertex list (mm).
+    X, Y = bounding box centroid
+    Z = minimum Z (bottom face of element — more natural for IFC)
+    Single-pass to avoid creating 3 sliced copies of a large list.
+    """
+    x0 = flat_verts[0]
+    y0 = flat_verts[1]
+    z0 = flat_verts[2]
+    xmin = xmax = x0
+    ymin = ymax = y0
+    zmin = z0
+    for i in range(3, len(flat_verts) - 2, 3):
+        x = flat_verts[i]
+        y = flat_verts[i + 1]
+        z = flat_verts[i + 2]
+        if x < xmin:
+            xmin = x
+        elif x > xmax:
+            xmax = x
+        if y < ymin:
+            ymin = y
+        elif y > ymax:
+            ymax = y
+        if z < zmin:
+            zmin = z
+    return (xmin + xmax) / 2.0, (ymin + ymax) / 2.0, zmin
+
+
+# Cache for shared IFC direction/point entities (keyed by ifc file id)
+_shared_entities: dict[int, dict] = {}
+
+
+def _get_shared(ifc):
+    """Return (or create) shared IfcDirection and IfcCartesianPoint entities for this file."""
+    fid = id(ifc)
+    if fid not in _shared_entities:
+        _shared_entities[fid] = {
+            "z_axis": ifc.createIfcDirection([0.0, 0.0, 1.0]),
+            "x_axis": ifc.createIfcDirection([1.0, 0.0, 0.0]),
+            "origin_0": ifc.createIfcCartesianPoint([0.0, 0.0, 0.0]),
+        }
+    return _shared_entities[fid]
+
+
+def _make_placement(ifc, x: float, y: float, z: float):
+    """Create an IfcLocalPlacement at absolute world coordinates (mm)."""
+    shared = _get_shared(ifc)
+    origin = ifc.createIfcCartesianPoint([round(x, 3), round(y, 3), round(z, 3)])
+    a2p    = ifc.createIfcAxis2Placement3D(origin, shared["z_axis"], shared["x_axis"])
+    return ifc.createIfcLocalPlacement(PlacementRelTo=None, RelativePlacement=a2p)
+
+
+# --------------------------------------------------------------------------- #
+# Main conversion
+# --------------------------------------------------------------------------- #
+
+def mesh_to_ifc(
+    ifc: ifcopenshell.file,
+    body_context,
+    obj: Base,
+    scale: float = 0.001,
+    material_manager=None,
+) -> tuple:
+    """
+    Convert a Speckle DataObject → (IfcShapeRepresentation, IfcLocalPlacement).
+    Creates one IfcPolygonalFaceSet per mesh so each can carry its own material style.
+    Returns (None, None) if no usable geometry is found.
+    """
+    meshes = get_display_meshes(obj)
+    if not meshes:
+        return None, None
+
+    # Parent object's applicationId — used as fallback for material lookup
+    # when inner meshes (e.g. from BrepX) don't have their own applicationId
+    obj_app_id = _get(obj, "applicationId")
+
+    obj_scale = _resolve_scale(obj, scale)
+
+    # ------------------------------------------------------------------ #
+    # Pass 1: unpack and scale vertices once per mesh, compute origin
+    #         incrementally without accumulating all vertices in memory.
+    # ------------------------------------------------------------------ #
+    mesh_cache = []   # [scaled_verts_list] or None per mesh
+    xmin = ymin = zmin = float("inf")
+    xmax = ymax = float("-inf")
+    has_verts = False
+
+    for mesh in meshes:
+        raw_verts = _get(mesh, "vertices") or []
+        verts = unwrap_chunks(raw_verts if isinstance(raw_verts, list) else list(raw_verts))
+        if not verts:
+            mesh_cache.append(None)
+            continue
+        ms = _resolve_scale(mesh, obj_scale)
+        scaled = [float(v) * ms for v in verts]
+        mesh_cache.append(scaled)
+        has_verts = True
+
+        # Update bounding box from this mesh's scaled vertices
+        for i in range(0, len(scaled) - 2, 3):
+            x, y, z = scaled[i], scaled[i + 1], scaled[i + 2]
+            if x < xmin: xmin = x
+            if x > xmax: xmax = x
+            if y < ymin: ymin = y
+            if y > ymax: ymax = y
+            if z < zmin: zmin = z
+
+    if not has_verts:
+        return None, None
+
+    ox = (xmin + xmax) / 2.0
+    oy = (ymin + ymax) / 2.0
+    oz = zmin
+
+    # ------------------------------------------------------------------ #
+    # Pass 2: one faceset per mesh — reuse cached verts, only unpack faces
+    # ------------------------------------------------------------------ #
+    geom_items = []
+
+    for mesh, scaled in zip(meshes, mesh_cache):
+        if scaled is None:
+            continue
+        raw_faces = _get(mesh, "faces") or []
+        faces_raw = unwrap_chunks(raw_faces if isinstance(raw_faces, list) else list(raw_faces))
+
+        if not faces_raw:
+            continue
+
+        try:
+            face_groups = decode_faces(faces_raw)
+        except Exception as e:
+            print(f"  Warning: Face decode error: {e}")
+            continue
+
+        # Offset pre-scaled vertices relative to origin (flat list, no tuples)
+        n = len(scaled)
+        verts_scaled = [0.0] * n
+        for vi in range(0, n, 3):
+            verts_scaled[vi]     = scaled[vi]     - ox
+            verts_scaled[vi + 1] = scaled[vi + 1] - oy
+            verts_scaled[vi + 2] = scaled[vi + 2] - oz
+
+        mesh_facesets = build_ifc_facesets(ifc, verts_scaled, face_groups)
+
+        if not mesh_facesets:
+            continue
+
+        # Apply material style to every faceset of this mesh
+        # Inner meshes (from BrepX) may lack applicationId — fall back to parent's
+        if material_manager:
+            mesh_app_id = _get(mesh, "applicationId") or obj_app_id
+            if mesh_app_id:
+                for fs in mesh_facesets:
+                    material_manager.apply_to_item(fs, str(mesh_app_id))
+
+        geom_items.extend(mesh_facesets)
+
+    if not geom_items:
+        return None, None
+
+    # ------------------------------------------------------------------ #
+    # Assemble IfcShapeRepresentation + IfcLocalPlacement
+    # ------------------------------------------------------------------ #
+    rep = ifc.createIfcShapeRepresentation(
+        ContextOfItems=body_context,
+        RepresentationIdentifier="Body",
+        RepresentationType="Tessellation",
+        Items=geom_items,
+    )
+    placement = _make_placement(ifc, ox, oy, oz)
+
+    return rep, placement
\ No newline at end of file
diff --git a/sample_files/instances.py b/sample_files/instances.py
new file mode 100644
index 0000000..23c298e
--- /dev/null
+++ b/sample_files/instances.py
@@ -0,0 +1,640 @@
+# =============================================================================
+# instances.py
+# Handles Speckle InstanceProxy objects from both:
+#
+# FORMAT A — Revit connector (our actual use case):
+#   _units      = "mm"
+#   transform   = 16 floats, row-major, translation in MM
+#   definitionId = 64-char uppercase hex hash (matches object id[:32] in tree)
+#   The definition object lives somewhere in the object tree.
+#
+# FORMAT B — speckleifc IFC→Speckle converter:
+#   units       = "m"
+#   transform   = 16 floats, row-major, translation in METRES
+#   definitionId = "DEFINITION:{meshAppId}"
+#   Definition geometry lives in root → Collection("definitionGeometry")
+#
+# We detect the format by the definitionId prefix.
+#
+# Performance: uses IfcRepresentationMap + IfcMappedItem so that all instances
+# sharing the same definition reference a single copy of the geometry.
+# =============================================================================
+
+import hashlib
+import math
+import struct
+import ifcopenshell.api
+from specklepy.objects.base import Base
+from utils.helpers import _get, MM_SCALES
+from utils.geometry import unwrap_chunks, decode_faces, build_ifc_facesets, _get_shared, _is_mesh
+from utils.curves import is_curve, build_curve_rep_map
+
+
+def is_instance(obj) -> bool:
+    """Returns True if this object is a Speckle InstanceProxy."""
+    return _get(obj, "transform") is not None and _get(obj, "definitionId") is not None
+
+
+def _is_ifc_format(definition_id: str) -> bool:
+    """True if this is speckleifc format (definitionId starts with 'DEFINITION:')."""
+    return definition_id.startswith("DEFINITION:")
+
+
+def build_definition_map(root: Base) -> dict:
+    """
+    Build a unified definition map that handles both formats.
+
+    Returns dict with keys:
+      "by_id"             : {obj_id_lower[:32] → object}   for Revit format
+      "by_app_id"         : {applicationId_lower → object}  for Revit format
+      "ifc_proxies"       : {"DEFINITION:xxx" → proxy}      for IFC format
+      "ifc_meshes"        : {meshAppId → Mesh}               for IFC format
+      "definition_sources": set of applicationId (lowercase) that are definition
+                            geometry sources — these should be skipped during export
+    """
+    by_id      = {}
+    by_app_id  = {}
+    ifc_proxies = {}
+    ifc_meshes  = {}
+    definition_sources = set()  # applicationIds used as definition geometry (skip during export)
+
+    # --- Walk entire tree for Revit format ---
+    _collect_all(root, by_id, by_app_id, depth=0)
+
+    # --- Extract speckleifc structures for IFC format ---
+    proxies_raw = _get(root, "instanceDefinitionProxies")
+    if proxies_raw:
+        for proxy in (proxies_raw if isinstance(proxies_raw, list) else [proxies_raw]):
+            app_id = _get(proxy, "applicationId")
+            if app_id:
+                ifc_proxies[app_id] = proxy            # original case (for IFC format)
+                ifc_proxies[app_id.lower()] = proxy    # lowercase (for Revit format)
+            # Collect all objects referenced by this proxy as definition sources
+            object_ids = _get(proxy, "objects") or []
+            for oid in (object_ids if isinstance(object_ids, list) else [object_ids]):
+                if oid:
+                    definition_sources.add(str(oid).lower())
+
+    elements = _get(root, "elements") or _get(root, "@elements") or []
+    for child in (elements if isinstance(elements, list) else []):
+        if (_get(child, "name") or "") == "definitionGeometry":
+            geom_elements = _get(child, "elements") or _get(child, "@elements") or []
+            for mesh in (geom_elements if isinstance(geom_elements, list) else []):
+                mesh_app_id = _get(mesh, "applicationId")
+                if mesh_app_id:
+                    ifc_meshes[mesh_app_id] = mesh
+
+    print(f"   Objects indexed by id:     {len(by_id)}")
+    print(f"   Objects indexed by appId:  {len(by_app_id)}")
+    print(f"   IFC definition proxies:    {len(ifc_proxies)}")
+    print(f"   IFC definition meshes:     {len(ifc_meshes)}")
+    print(f"   Definition sources:        {len(definition_sources)}")
+
+    return {
+        "by_id":              by_id,
+        "by_app_id":          by_app_id,
+        "ifc_proxies":        ifc_proxies,
+        "ifc_meshes":         ifc_meshes,
+        "definition_sources": definition_sources,
+    }
+
+
+def _collect_all(obj, by_id: dict, by_app_id: dict, depth: int):
+    if obj is None or depth > 25:
+        return
+
+    obj_id = _get(obj, "id")
+    if obj_id and isinstance(obj_id, str):
+        key = obj_id.lower()
+        by_id[key] = obj
+        # Also store truncated — definitionId (64 chars) matches id (32 chars)
+        if len(key) == 32:
+            by_id[key] = obj
+        elif len(key) > 32:
+            by_id[key[:32]] = obj
+
+    app_id = _get(obj, "applicationId")
+    if app_id and isinstance(app_id, str):
+        by_app_id[app_id.lower()] = obj
+
+    for key in ["elements", "@elements", "_elements",
+                "displayValue", "@displayValue", "_displayValue",
+                "objects", "@objects", "definition", "@definition"]:
+        try:
+            children = obj[key]
+            if children is None:
+                continue
+            if not isinstance(children, list):
+                children = [children]
+            for child in children:
+                _collect_all(child, by_id, by_app_id, depth + 1)
+        except Exception:
+            continue
+
+
+def _get_definition_source_object(definition_id: str, definition_map: dict):
+    """Resolve the first source object referenced by a definition proxy."""
+    ifc_proxies = definition_map.get("ifc_proxies", {})
+    proxy = ifc_proxies.get(definition_id) or ifc_proxies.get(definition_id.lower())
+    if proxy is None:
+        return None
+    object_ids = _get(proxy, "objects") or []
+    if not isinstance(object_ids, list):
+        object_ids = list(object_ids)
+    if not object_ids:
+        return None
+    by_app_id = definition_map.get("by_app_id", {})
+    return by_app_id.get(str(object_ids[0]).lower())
+
+
+def _get_revit_meshes(definition_id: str, definition_map: dict) -> tuple:
+    """
+    Revit format:
+      definitionId (64-char hex) → InstanceDefinitionProxy.applicationId
+      proxy.objects[0] is a UUID applicationId → find mesh by applicationId
+
+    Returns (meshes, app_ids) where app_ids are all applicationIds encountered
+    in the resolution chain (definition objects, geometry objects) for material fallback.
+    """
+    from utils.geometry import get_display_meshes
+
+    # Step 1: find the InstanceDefinitionProxy by its applicationId (case-insensitive)
+    ifc_proxies = definition_map.get("ifc_proxies", {})
+    proxy = ifc_proxies.get(definition_id) or ifc_proxies.get(definition_id.lower())
+    if proxy is None:
+        return [], []
+
+    # Step 2: get the mesh applicationIds from proxy.objects
+    object_ids = _get(proxy, "objects") or []
+    if not isinstance(object_ids, list):
+        object_ids = list(object_ids)
+
+    # Step 3: look up each mesh by applicationId, collecting all encountered app IDs
+    by_app_id = definition_map.get("by_app_id", {})
+    meshes = []
+    encountered_app_ids = []
+    for oid in object_ids:
+        obj = by_app_id.get(str(oid).lower())
+        if obj is not None:
+            # Collect this object's applicationId
+            obj_aid = _get(obj, "applicationId")
+            if obj_aid:
+                encountered_app_ids.append(str(obj_aid))
+            # Also collect applicationIds from displayValue items (BrepX, etc.)
+            for key in ["displayValue", "@displayValue", "_displayValue"]:
+                display = _get(obj, key)
+                if display:
+                    items = display if isinstance(display, list) else [display]
+                    for item in items:
+                        item_aid = _get(item, "applicationId")
+                        if item_aid:
+                            encountered_app_ids.append(str(item_aid))
+                    break
+            # The found object may itself be a mesh, or contain displayValue meshes
+            found_meshes = get_display_meshes(obj)
+            if found_meshes:
+                meshes.extend(found_meshes)
+            elif _is_mesh(obj):
+                # Object itself is a mesh (no displayValue wrapping)
+                meshes.append(obj)
+    return meshes, encountered_app_ids
+
+
+def _get_ifc_meshes(definition_id: str, definition_map: dict) -> tuple:
+    """
+    IFC format: definitionId = "DEFINITION:224058_mat0"
+    Look up proxy → objects list → meshes from ifc_meshes dict.
+    Returns (meshes, []) — no extra app_ids needed, mesh applicationIds match directly.
+    """
+    ifc_proxies = definition_map.get("ifc_proxies", {})
+    ifc_meshes  = definition_map.get("ifc_meshes", {})
+
+    proxy = ifc_proxies.get(definition_id)
+    if proxy is None:
+        return [], []
+
+    object_ids = _get(proxy, "objects") or []
+    result = []
+    for oid in (object_ids if isinstance(object_ids, list) else [object_ids]):
+        mesh = ifc_meshes.get(str(oid))
+        if mesh is not None:
+            result.append(mesh)
+    return result, []
+
+
+def _resolve_instance_scale(obj, stream_scale: float) -> float:
+    """
+    Resolve scale for the transform translation.
+    Tries bracket access for '_units' (Revit uses underscore).
+    IFC format instances have units="m" → scale=1.0 (no scaling).
+    """
+    for key in ["units", "_units"]:
+        try:
+            units = obj[key]
+            if units and isinstance(units, str):
+                s = MM_SCALES.get(units.lower().strip())
+                if s is not None:
+                    return s
+        except Exception:
+            pass
+    return stream_scale
+
+
+# Stats
+_stats = {"found": 0, "not_found": 0}
+
+# Cache: mesh id → (verts_scaled, face_groups) to avoid re-unpacking
+# AND re-scaling the same definition mesh across many instances that share it.
+_mesh_data_cache: dict = {}
+
+# Cache: definition_id → IfcRepresentationMap (or None if no geometry)
+# All instances sharing the same definition reuse one geometry copy.
+_rep_map_cache: dict = {}
+
+# Cache: geometry content hash → IfcRepresentationMap
+# Enables sharing across different definitionIds that have identical geometry.
+_geometry_hash_cache: dict = {}
+
+# Shared identity placement for all instances (keyed by ifc file id)
+_identity_placement_cache: dict[int, object] = {}
+
+
+# --------------------------------------------------------------------------- #
+# Geometry content hashing
+# --------------------------------------------------------------------------- #
+
+def _hash_mesh_data(mesh_data_list: list, material_key: str = "") -> str:
+    """Compute a content hash from mesh geometry data for deduplication.
+
+    mesh_data_list: list of (verts_local, face_groups) tuples
+    material_key:   string identifying the material (included in hash)
+    Returns: hex digest string
+    """
+    h = hashlib.md5(usedforsecurity=False)
+    for verts_local, face_groups in mesh_data_list:
+        # Hash rounded vertices as packed floats (faster than str conversion)
+        for i in range(0, len(verts_local), 3):
+            h.update(struct.pack("3f",
+                round(verts_local[i], 3),
+                round(verts_local[i+1], 3),
+                round(verts_local[i+2], 3),
+            ))
+        # Hash face indices
+        for face in face_groups:
+            h.update(struct.pack(f"{len(face)}i", *face))
+        # Separator between meshes
+        h.update(b"|")
+    if material_key:
+        h.update(material_key.encode())
+    return h.hexdigest()
+
+
+# --------------------------------------------------------------------------- #
+# IfcRepresentationMap builder — geometry created once per definition
+# --------------------------------------------------------------------------- #
+
+def _collect_mesh_data(meshes: list, ifc_format: bool) -> list:
+    """Unpack, scale, and cache mesh vertex/face data.
+
+    Returns list of (mesh_obj, verts_local, face_groups) tuples.
+    """
+    result = []
+    for mesh in meshes:
+        mesh_id = _get(mesh, "id") or _get(mesh, "applicationId")
+        if mesh_id and mesh_id in _mesh_data_cache:
+            verts_local, face_groups = _mesh_data_cache[mesh_id]
+        else:
+            raw_verts = _get(mesh, "vertices") or []
+            raw_faces = _get(mesh, "faces") or []
+            verts     = unwrap_chunks(raw_verts if isinstance(raw_verts, list) else list(raw_verts))
+            faces_raw = unwrap_chunks(raw_faces if isinstance(raw_faces, list) else list(raw_faces))
+            if not verts or not faces_raw:
+                continue
+
+            mesh_units = _get(mesh, "units") or _get(mesh, "_units") or ("m" if ifc_format else "mm")
+            ms = MM_SCALES.get(mesh_units.lower().strip(), 1.0)
+
+            try:
+                face_groups = decode_faces(faces_raw)
+            except Exception as e:
+                print(f"  Warning: Instance face decode: {e}")
+                continue
+
+            verts_local = [float(v) * ms for v in verts]
+
+            if mesh_id:
+                _mesh_data_cache[mesh_id] = (verts_local, face_groups)
+
+        result.append((mesh, verts_local, face_groups))
+    return result
+
+
+def _resolve_material_key(meshes_data: list, material_manager, fallback_app_ids, definition_id) -> str:
+    """Build a material cache key string for geometry hashing."""
+    if not material_manager:
+        return ""
+    parts = []
+    for mesh, _, _ in meshes_data:
+        mesh_app_id = _get(mesh, "applicationId")
+        style = material_manager.get_style_with_fallbacks(
+            primary_app_id=str(mesh_app_id) if mesh_app_id else None,
+            fallback_app_ids=fallback_app_ids,
+            definition_id=definition_id,
+        )
+        parts.append(str(id(style)) if style else "")
+    return "|".join(parts)
+
+
+def _build_rep_map(ifc, body_context, meshes: list, ifc_format: bool,
+                   material_manager=None, fallback_app_ids: list = None,
+                   definition_id: str = None):
+    """
+    Build an IfcRepresentationMap from definition meshes.
+    Uses content-based hashing to reuse identical geometry across different
+    definitionIds. Returns IfcRepresentationMap or None if no valid geometry.
+    """
+    # Step 1: Collect and cache raw mesh data (no IFC entities created yet)
+    meshes_data = _collect_mesh_data(meshes, ifc_format)
+    if not meshes_data:
+        return None
+
+    # Step 2: Compute content hash to check for identical geometry
+    mat_key = _resolve_material_key(meshes_data, material_manager, fallback_app_ids, definition_id)
+    geom_hash = _hash_mesh_data(
+        [(verts, faces) for _, verts, faces in meshes_data],
+        material_key=mat_key,
+    )
+
+    if geom_hash in _geometry_hash_cache:
+        return _geometry_hash_cache[geom_hash]
+
+    # Step 3: No match — build IFC geometry entities
+    geom_items = []
+
+    for mesh, verts_local, face_groups in meshes_data:
+        mesh_facesets = build_ifc_facesets(ifc, verts_local, face_groups)
+        if not mesh_facesets:
+            continue
+
+        if material_manager:
+            mesh_app_id = _get(mesh, "applicationId")
+            style = material_manager.get_style_with_fallbacks(
+                primary_app_id=str(mesh_app_id) if mesh_app_id else None,
+                fallback_app_ids=fallback_app_ids,
+                definition_id=definition_id,
+            )
+            if style:
+                for fs in mesh_facesets:
+                    try:
+                        ifcopenshell.api.run(
+                            "style.assign_item_style", ifc,
+                            item=fs, style=style,
+                        )
+                        material_manager._apply_count += 1
+                    except Exception:
+                        pass
+
+        geom_items.extend(mesh_facesets)
+
+    if not geom_items:
+        _geometry_hash_cache[geom_hash] = None
+        return None
+
+    shared = _get_shared(ifc)
+    a2p    = ifc.createIfcAxis2Placement3D(shared["origin_0"], None, None)
+
+    mapped_rep = ifc.createIfcShapeRepresentation(
+        ContextOfItems=body_context,
+        RepresentationIdentifier="Body",
+        RepresentationType="Tessellation",
+        Items=geom_items,
+    )
+
+    rep_map = ifc.createIfcRepresentationMap(a2p, mapped_rep)
+    _geometry_hash_cache[geom_hash] = rep_map
+    return rep_map
+
+
+# --------------------------------------------------------------------------- #
+# Transform → IfcCartesianTransformationOperator3D
+# --------------------------------------------------------------------------- #
+
+def _vec_magnitude(x, y, z):
+    return math.sqrt(x*x + y*y + z*z)
+
+
+# Cache: rounded direction tuple → IfcDirection entity (keyed by ifc file id)
+_direction_cache: dict[int, dict] = {}
+
+def _get_or_create_direction(ifc, dx, dy, dz):
+    """Return a cached IfcDirection or create and cache a new one."""
+    fid = id(ifc)
+    if fid not in _direction_cache:
+        _direction_cache[fid] = {}
+    cache = _direction_cache[fid]
+    # Round to 6 decimals — sufficient for unit vectors
+    key = (round(dx, 6), round(dy, 6), round(dz, 6))
+    if key not in cache:
+        cache[key] = ifc.createIfcDirection([key[0], key[1], key[2]])
+    return cache[key]
+
+
+def _make_transform_operator(ifc, t: list, ts: float):
+    """
+    Convert a row-major 4x4 matrix + translation scale into an
+    IfcCartesianTransformationOperator3DnonUniform.
+
+    t:  16 floats, row-major [r00,r01,r02,tx, r10,r11,r12,ty, r20,r21,r22,tz, 0,0,0,1]
+    ts: scale factor for translation components (e.g. 1000.0 for m→mm)
+
+    IfcCartesianTransformationOperator axes represent the COLUMNS of M:
+      Axis1 = column 0 = where local X maps → (t[0], t[4], t[8])
+      Axis2 = column 1 = where local Y maps → (t[1], t[5], t[9])
+      Axis3 = column 2 = where local Z maps → (t[2], t[6], t[10])
+
+    Always uses the non-uniform variant with explicit Axis3 to ensure
+    correct orientation for all transform types (mirrors, non-orthogonal, etc.).
+
+    Returns the IFC entity, or None if the transform is degenerate.
+    """
+    # Extract COLUMNS of the 3x3 rotation/scale sub-matrix
+    ax1 = (float(t[0]), float(t[4]), float(t[8]))
+    ax2 = (float(t[1]), float(t[5]), float(t[9]))
+    ax3 = (float(t[2]), float(t[6]), float(t[10]))
+
+    s1 = _vec_magnitude(*ax1)
+    s2 = _vec_magnitude(*ax2)
+    s3 = _vec_magnitude(*ax3)
+
+    if s1 < 1e-10 or s2 < 1e-10 or s3 < 1e-10:
+        return None  # degenerate transform
+
+    # Normalized direction vectors — reuse cached IfcDirection entities
+    d1 = _get_or_create_direction(ifc, ax1[0]/s1, ax1[1]/s1, ax1[2]/s1)
+    d2 = _get_or_create_direction(ifc, ax2[0]/s2, ax2[1]/s2, ax2[2]/s2)
+    d3 = _get_or_create_direction(ifc, ax3[0]/s3, ax3[1]/s3, ax3[2]/s3)
+
+    # Translation, scaled and rounded to mm
+    tx = round(float(t[3])  * ts, 3)
+    ty = round(float(t[7])  * ts, 3)
+    tz = round(float(t[11]) * ts, 3)
+    origin = ifc.createIfcCartesianPoint([tx, ty, tz])
+
+    # Round scales for cleaner output
+    s1 = round(s1, 6)
+    s2 = round(s2, 6)
+    s3 = round(s3, 6)
+
+    return ifc.createIfcCartesianTransformationOperator3DnonUniform(
+        d1,      # Axis1
+        d2,      # Axis2
+        origin,  # LocalOrigin
+        s1,      # Scale
+        d3,      # Axis3 (explicit — never derived)
+        s2,      # Scale2
+        s3,      # Scale3
+    )
+
+
+# --------------------------------------------------------------------------- #
+# Main conversion — IfcMappedItem approach
+# --------------------------------------------------------------------------- #
+
+def instance_to_ifc(ifc, body_context, obj: Base, definition_map: dict,
+                    scale: float = 1.0, material_manager=None):
+    """
+    Convert a Speckle InstanceProxy → (IfcShapeRepresentation, IfcLocalPlacement).
+
+    Strategy: create geometry once per definition as an IfcRepresentationMap,
+    then reference it via IfcMappedItem + IfcCartesianTransformationOperator3D
+    for each instance. This avoids duplicating geometry across instances.
+    """
+    transform_raw = _get(obj, "transform")
+    if not transform_raw:
+        return None, None
+    t = list(transform_raw)
+    if len(t) != 16:
+        return None, None
+
+    definition_id = _get(obj, "definitionId") or ""
+    ifc_format    = _is_ifc_format(definition_id)
+
+    # Translation scale: IFC format transform is in metres → convert to mm
+    # Revit format transform is already in mm (same as IFC file units)
+    ts = 1000.0 if ifc_format else _resolve_instance_scale(obj, scale)
+
+    # Identity placement (transform is encoded in the MappedItem) — shared across all instances
+    fid = id(ifc)
+    if fid not in _identity_placement_cache:
+        shared = _get_shared(ifc)
+        a2p = ifc.createIfcAxis2Placement3D(shared["origin_0"], None, None)
+        _identity_placement_cache[fid] = ifc.createIfcLocalPlacement(PlacementRelTo=None, RelativePlacement=a2p)
+    placement = _identity_placement_cache[fid]
+
+    # --- Get or build IfcRepresentationMap (cached per definition_id) ---
+    if definition_id not in _rep_map_cache:
+        if ifc_format:
+            meshes, extra_app_ids = _get_ifc_meshes(definition_id, definition_map)
+        else:
+            meshes, extra_app_ids = _get_revit_meshes(definition_id, definition_map)
+
+        # Build fallback app_id list: instance's own + definition chain IDs
+        instance_app_id = _get(obj, "applicationId")
+        fallback_ids = []
+        if instance_app_id:
+            fallback_ids.append(str(instance_app_id))
+        fallback_ids.extend(extra_app_ids)
+
+        rep_map_result = None
+        if meshes:
+            rep_map_result = _build_rep_map(
+                ifc, body_context, meshes, ifc_format, material_manager,
+                fallback_app_ids=fallback_ids,
+                definition_id=definition_id,
+            )
+
+        # If no mesh geometry produced, try curve geometry from the definition object
+        if rep_map_result is None:
+            curve_obj = _get_definition_source_object(definition_id, definition_map)
+            if curve_obj and is_curve(curve_obj):
+                curve_scale = _resolve_instance_scale(curve_obj, 1.0)
+                rep_map_result = build_curve_rep_map(
+                    ifc, body_context, curve_obj, scale=curve_scale,
+                    material_manager=material_manager,
+                    fallback_app_ids=fallback_ids,
+                    definition_id=definition_id,
+                )
+
+        _rep_map_cache[definition_id] = rep_map_result
+        if rep_map_result is not None:
+            _stats["found"] += 1
+        else:
+            _stats["not_found"] += 1
+    else:
+        # Track stats even for cached definitions
+        if _rep_map_cache[definition_id] is not None:
+            _stats["found"] += 1
+        else:
+            _stats["not_found"] += 1
+
+    rep_map = _rep_map_cache[definition_id]
+    if rep_map is None:
+        return None, placement
+
+    # --- Build transform operator from instance's 4x4 matrix ---
+    transform_op = _make_transform_operator(ifc, t, ts)
+    if transform_op is None:
+        return None, placement
+
+    # --- Create IfcMappedItem referencing the shared geometry ---
+    mapped_item = ifc.createIfcMappedItem(rep_map, transform_op)
+
+    rep = ifc.createIfcShapeRepresentation(
+        ContextOfItems=body_context,
+        RepresentationIdentifier="Body",
+        RepresentationType="MappedRepresentation",
+        Items=[mapped_item],
+    )
+    return rep, placement
+
+
+def get_definition_object(obj: Base, definition_map: dict):
+    """
+    Resolve the definition's source object for an InstanceProxy.
+    Returns the first object referenced by the definition proxy, which
+    carries the proper category/type info. Returns None if not found.
+    """
+    definition_id = _get(obj, "definitionId") or ""
+    if not definition_id:
+        return None
+    return _get_definition_source_object(definition_id, definition_map)
+
+
+def is_definition_source(obj, definition_map: dict) -> bool:
+    """Return True if this object is a definition geometry source (should not be exported standalone)."""
+    app_id = _get(obj, "applicationId")
+    if not app_id:
+        return False
+    return str(app_id).lower() in definition_map.get("definition_sources", set())
+
+
+def print_instance_stats():
+    total = _stats["found"] + _stats["not_found"]
+    print(f"  Instance resolution: {_stats['found']}/{total} definitions found")
+    if _stats["not_found"] > 0:
+        print(f"  Warning: {_stats['not_found']} instances had no definition geometry")
+    unique_defs = len(_rep_map_cache)
+    unique_geom = len([v for v in _geometry_hash_cache.values() if v is not None])
+    if unique_defs > unique_geom:
+        print(f"  Geometry dedup: {unique_defs} definitions -> {unique_geom} unique geometries")
+
+
+def reset_caches():
+    """Reset module-level caches (call at start of each export run)."""
+    _mesh_data_cache.clear()
+    _rep_map_cache.clear()
+    _geometry_hash_cache.clear()
+    _identity_placement_cache.clear()
+    _direction_cache.clear()
+    _stats["found"] = 0
+    _stats["not_found"] = 0
diff --git a/utils/curves.py b/utils/curves.py
index a31fa71..c9168a6 100644
--- a/utils/curves.py
+++ b/utils/curves.py
@@ -15,7 +15,7 @@
 import ifcopenshell
 import ifcopenshell.api
 from specklepy.objects.base import Base
-from utils.helpers import _get, MM_SCALES
+from utils.helpers import _get, resolve_scale as _resolve_scale
 from utils.geometry import _get_shared, _make_placement
 
 
@@ -29,14 +29,6 @@ def is_curve(obj) -> bool:
     return any(ct in speckle_type for ct in _CURVE_TYPES)
 
 
-def _resolve_scale(obj, fallback: float) -> float:
-    """Resolve unit scale for a curve object."""
-    units = _get(obj, "units")
-    if units and isinstance(units, str):
-        return MM_SCALES.get(units.lower().strip(), fallback)
-    return fallback
-
-
 def _point_coords(pt, scale: float) -> tuple:
     """Extract (x, y, z) from a Speckle Point, scaled to mm."""
     x = float(_get(pt, "x") or 0) * scale
@@ -183,7 +175,9 @@ def build_ifc_curve(ifc, points: list, segments: list):
     if not points or not segments:
         return None
 
-    point_list = ifc.createIfcCartesianPointList3D(points)
+    # Round coordinates for smaller IFC file size (0.001mm precision)
+    rounded = [[round(p[0], 3), round(p[1], 3), round(p[2], 3)] for p in points]
+    point_list = ifc.createIfcCartesianPointList3D(rounded)
 
     ifc_segments = []
     for seg_type, indices in segments:
diff --git a/utils/geometry.py b/utils/geometry.py
index eb1644d..e035724 100644
--- a/utils/geometry.py
+++ b/utils/geometry.py
@@ -13,7 +13,7 @@
 
 import ifcopenshell
 from specklepy.objects.base import Base
-from utils.helpers import _get, MM_SCALES as _UNIT_SCALES
+from utils.helpers import _get, resolve_scale as _resolve_scale
 
 
 # --------------------------------------------------------------------------- #
@@ -76,6 +76,13 @@ def build_ifc_facesets(ifc, verts_scaled: list, face_groups: list) -> list:
     if not valid_faces or not deduped_verts:
         return []
 
+    # Round vertex coordinates to reduce IFC text file size
+    # 3 decimal places = 0.001mm precision (more than sufficient)
+    for v in deduped_verts:
+        v[0] = round(v[0], 3)
+        v[1] = round(v[1], 3)
+        v[2] = round(v[2], 3)
+
     # Build IFC entities
     try:
         point_list = ifc.createIfcCartesianPointList3D(deduped_verts)
@@ -127,14 +134,6 @@ def unwrap_chunks(raw) -> list:
     return result
 
 
-def _resolve_scale(obj, stream_scale: float) -> float:
-    """Resolve unit scale: obj.units → stream fallback."""
-    units = _get(obj, "units")
-    if units and isinstance(units, str):
-        return _UNIT_SCALES.get(units.lower().strip(), stream_scale)
-    return stream_scale
-
-
 # --------------------------------------------------------------------------- #
 # Mesh extraction
 # --------------------------------------------------------------------------- #
@@ -264,36 +263,6 @@ def decode_faces(faces_raw: list) -> list:
 # Bounding box + placement
 # --------------------------------------------------------------------------- #
 
-def compute_origin(flat_verts: list) -> tuple:
-    """
-    Compute placement origin from scaled vertex list (mm).
-    X, Y = bounding box centroid
-    Z = minimum Z (bottom face of element — more natural for IFC)
-    Single-pass to avoid creating 3 sliced copies of a large list.
-    """
-    x0 = flat_verts[0]
-    y0 = flat_verts[1]
-    z0 = flat_verts[2]
-    xmin = xmax = x0
-    ymin = ymax = y0
-    zmin = z0
-    for i in range(3, len(flat_verts) - 2, 3):
-        x = flat_verts[i]
-        y = flat_verts[i + 1]
-        z = flat_verts[i + 2]
-        if x < xmin:
-            xmin = x
-        elif x > xmax:
-            xmax = x
-        if y < ymin:
-            ymin = y
-        elif y > ymax:
-            ymax = y
-        if z < zmin:
-            zmin = z
-    return (xmin + xmax) / 2.0, (ymin + ymax) / 2.0, zmin
-
-
 # Cache for shared IFC direction/point entities (keyed by ifc file id)
 _shared_entities: dict[int, dict] = {}
 
@@ -313,7 +282,7 @@ def _get_shared(ifc):
 def _make_placement(ifc, x: float, y: float, z: float):
     """Create an IfcLocalPlacement at absolute world coordinates (metres)."""
     shared = _get_shared(ifc)
-    origin = ifc.createIfcCartesianPoint([x, y, z])
+    origin = ifc.createIfcCartesianPoint([round(x, 3), round(y, 3), round(z, 3)])
     a2p    = ifc.createIfcAxis2Placement3D(origin, shared["z_axis"], shared["x_axis"])
     return ifc.createIfcLocalPlacement(PlacementRelTo=None, RelativePlacement=a2p)
 
diff --git a/utils/helpers.py b/utils/helpers.py
index 94ddbd3..33514c2 100644
--- a/utils/helpers.py
+++ b/utils/helpers.py
@@ -36,3 +36,11 @@ MM_SCALES = {
     "ft": 304.8,  "foot": 304.8,        "feet": 304.8,
     "in": 25.4,   "inch": 25.4,         "inches": 25.4,
 }
+
+
+def resolve_scale(obj, fallback: float) -> float:
+    """Resolve unit scale: obj.units → fallback."""
+    units = _get(obj, "units")
+    if units and isinstance(units, str):
+        return MM_SCALES.get(units.lower().strip(), fallback)
+    return fallback
diff --git a/utils/instances.py b/utils/instances.py
index 7a97f50..23c298e 100644
--- a/utils/instances.py
+++ b/utils/instances.py
@@ -20,7 +20,9 @@
 # sharing the same definition reference a single copy of the geometry.
 # =============================================================================
 
+import hashlib
 import math
+import struct
 import ifcopenshell.api
 from specklepy.objects.base import Base
 from utils.helpers import _get, MM_SCALES
@@ -249,24 +251,54 @@ _mesh_data_cache: dict = {}
 # All instances sharing the same definition reuse one geometry copy.
 _rep_map_cache: dict = {}
 
+# Cache: geometry content hash → IfcRepresentationMap
+# Enables sharing across different definitionIds that have identical geometry.
+_geometry_hash_cache: dict = {}
+
 # Shared identity placement for all instances (keyed by ifc file id)
 _identity_placement_cache: dict[int, object] = {}
 
 
+# --------------------------------------------------------------------------- #
+# Geometry content hashing
+# --------------------------------------------------------------------------- #
+
+def _hash_mesh_data(mesh_data_list: list, material_key: str = "") -> str:
+    """Compute a content hash from mesh geometry data for deduplication.
+
+    mesh_data_list: list of (verts_local, face_groups) tuples
+    material_key:   string identifying the material (included in hash)
+    Returns: hex digest string
+    """
+    h = hashlib.md5(usedforsecurity=False)
+    for verts_local, face_groups in mesh_data_list:
+        # Hash rounded vertices as packed floats (faster than str conversion)
+        for i in range(0, len(verts_local), 3):
+            h.update(struct.pack("3f",
+                round(verts_local[i], 3),
+                round(verts_local[i+1], 3),
+                round(verts_local[i+2], 3),
+            ))
+        # Hash face indices
+        for face in face_groups:
+            h.update(struct.pack(f"{len(face)}i", *face))
+        # Separator between meshes
+        h.update(b"|")
+    if material_key:
+        h.update(material_key.encode())
+    return h.hexdigest()
+
+
 # --------------------------------------------------------------------------- #
 # IfcRepresentationMap builder — geometry created once per definition
 # --------------------------------------------------------------------------- #
 
-def _build_rep_map(ifc, body_context, meshes: list, ifc_format: bool,
-                   material_manager=None, fallback_app_ids: list = None,
-                   definition_id: str = None):
-    """
-    Build an IfcRepresentationMap from definition meshes.
-    Geometry is in local coordinates (mm, no instance transform applied).
-    Returns IfcRepresentationMap or None if no valid geometry.
-    """
-    geom_items = []
+def _collect_mesh_data(meshes: list, ifc_format: bool) -> list:
+    """Unpack, scale, and cache mesh vertex/face data.
 
+    Returns list of (mesh_obj, verts_local, face_groups) tuples.
+    """
+    result = []
     for mesh in meshes:
         mesh_id = _get(mesh, "id") or _get(mesh, "applicationId")
         if mesh_id and mesh_id in _mesh_data_cache:
@@ -288,19 +320,62 @@ def _build_rep_map(ifc, body_context, meshes: list, ifc_format: bool,
                 print(f"  Warning: Instance face decode: {e}")
                 continue
 
-            # Scale vertices once and cache the result
             verts_local = [float(v) * ms for v in verts]
 
             if mesh_id:
                 _mesh_data_cache[mesh_id] = (verts_local, face_groups)
 
-        mesh_facesets = build_ifc_facesets(ifc, verts_local, face_groups)
+        result.append((mesh, verts_local, face_groups))
+    return result
 
+
+def _resolve_material_key(meshes_data: list, material_manager, fallback_app_ids, definition_id) -> str:
+    """Build a material cache key string for geometry hashing."""
+    if not material_manager:
+        return ""
+    parts = []
+    for mesh, _, _ in meshes_data:
+        mesh_app_id = _get(mesh, "applicationId")
+        style = material_manager.get_style_with_fallbacks(
+            primary_app_id=str(mesh_app_id) if mesh_app_id else None,
+            fallback_app_ids=fallback_app_ids,
+            definition_id=definition_id,
+        )
+        parts.append(str(id(style)) if style else "")
+    return "|".join(parts)
+
+
+def _build_rep_map(ifc, body_context, meshes: list, ifc_format: bool,
+                   material_manager=None, fallback_app_ids: list = None,
+                   definition_id: str = None):
+    """
+    Build an IfcRepresentationMap from definition meshes.
+    Uses content-based hashing to reuse identical geometry across different
+    definitionIds. Returns IfcRepresentationMap or None if no valid geometry.
+    """
+    # Step 1: Collect and cache raw mesh data (no IFC entities created yet)
+    meshes_data = _collect_mesh_data(meshes, ifc_format)
+    if not meshes_data:
+        return None
+
+    # Step 2: Compute content hash to check for identical geometry
+    mat_key = _resolve_material_key(meshes_data, material_manager, fallback_app_ids, definition_id)
+    geom_hash = _hash_mesh_data(
+        [(verts, faces) for _, verts, faces in meshes_data],
+        material_key=mat_key,
+    )
+
+    if geom_hash in _geometry_hash_cache:
+        return _geometry_hash_cache[geom_hash]
+
+    # Step 3: No match — build IFC geometry entities
+    geom_items = []
+
+    for mesh, verts_local, face_groups in meshes_data:
+        mesh_facesets = build_ifc_facesets(ifc, verts_local, face_groups)
         if not mesh_facesets:
             continue
 
-        # Apply material style to each faceset
-        # Try: mesh applicationId → fallback IDs → definitionId mapping
         if material_manager:
             mesh_app_id = _get(mesh, "applicationId")
             style = material_manager.get_style_with_fallbacks(
@@ -322,13 +397,12 @@ def _build_rep_map(ifc, body_context, meshes: list, ifc_format: bool,
         geom_items.extend(mesh_facesets)
 
     if not geom_items:
+        _geometry_hash_cache[geom_hash] = None
         return None
 
-    # Mapping origin = identity (local coords origin) — reuse shared origin
     shared = _get_shared(ifc)
     a2p    = ifc.createIfcAxis2Placement3D(shared["origin_0"], None, None)
 
-    # The mapped representation holds the actual geometry
     mapped_rep = ifc.createIfcShapeRepresentation(
         ContextOfItems=body_context,
         RepresentationIdentifier="Body",
@@ -336,7 +410,9 @@ def _build_rep_map(ifc, body_context, meshes: list, ifc_format: bool,
         Items=geom_items,
     )
 
-    return ifc.createIfcRepresentationMap(a2p, mapped_rep)
+    rep_map = ifc.createIfcRepresentationMap(a2p, mapped_rep)
+    _geometry_hash_cache[geom_hash] = rep_map
+    return rep_map
 
 
 # --------------------------------------------------------------------------- #
@@ -347,6 +423,22 @@ def _vec_magnitude(x, y, z):
     return math.sqrt(x*x + y*y + z*z)
 
 
+# Cache: rounded direction tuple → IfcDirection entity (keyed by ifc file id)
+_direction_cache: dict[int, dict] = {}
+
+def _get_or_create_direction(ifc, dx, dy, dz):
+    """Return a cached IfcDirection or create and cache a new one."""
+    fid = id(ifc)
+    if fid not in _direction_cache:
+        _direction_cache[fid] = {}
+    cache = _direction_cache[fid]
+    # Round to 6 decimals — sufficient for unit vectors
+    key = (round(dx, 6), round(dy, 6), round(dz, 6))
+    if key not in cache:
+        cache[key] = ifc.createIfcDirection([key[0], key[1], key[2]])
+    return cache[key]
+
+
 def _make_transform_operator(ifc, t: list, ts: float):
     """
     Convert a row-major 4x4 matrix + translation scale into an
@@ -355,22 +447,20 @@ def _make_transform_operator(ifc, t: list, ts: float):
     t:  16 floats, row-major [r00,r01,r02,tx, r10,r11,r12,ty, r20,r21,r22,tz, 0,0,0,1]
     ts: scale factor for translation components (e.g. 1000.0 for m→mm)
 
-    The matrix acts as: p' = M * p + translation, where M rows are:
-      row0 = (t[0], t[1], t[2])
-      row1 = (t[4], t[5], t[6])
-      row2 = (t[8], t[9], t[10])
-
     IfcCartesianTransformationOperator axes represent the COLUMNS of M:
       Axis1 = column 0 = where local X maps → (t[0], t[4], t[8])
       Axis2 = column 1 = where local Y maps → (t[1], t[5], t[9])
       Axis3 = column 2 = where local Z maps → (t[2], t[6], t[10])
 
+    Always uses the non-uniform variant with explicit Axis3 to ensure
+    correct orientation for all transform types (mirrors, non-orthogonal, etc.).
+
     Returns the IFC entity, or None if the transform is degenerate.
     """
     # Extract COLUMNS of the 3x3 rotation/scale sub-matrix
-    ax1 = (float(t[0]), float(t[4]), float(t[8]))    # column 0: X-axis direction
-    ax2 = (float(t[1]), float(t[5]), float(t[9]))    # column 1: Y-axis direction
-    ax3 = (float(t[2]), float(t[6]), float(t[10]))   # column 2: Z-axis direction
+    ax1 = (float(t[0]), float(t[4]), float(t[8]))
+    ax2 = (float(t[1]), float(t[5]), float(t[9]))
+    ax3 = (float(t[2]), float(t[6]), float(t[10]))
 
     s1 = _vec_magnitude(*ax1)
     s2 = _vec_magnitude(*ax2)
@@ -379,24 +469,28 @@ def _make_transform_operator(ifc, t: list, ts: float):
     if s1 < 1e-10 or s2 < 1e-10 or s3 < 1e-10:
         return None  # degenerate transform
 
-    # Normalized direction vectors
-    d1 = ifc.createIfcDirection([ax1[0]/s1, ax1[1]/s1, ax1[2]/s1])
-    d2 = ifc.createIfcDirection([ax2[0]/s2, ax2[1]/s2, ax2[2]/s2])
-    d3 = ifc.createIfcDirection([ax3[0]/s3, ax3[1]/s3, ax3[2]/s3])
+    # Normalized direction vectors — reuse cached IfcDirection entities
+    d1 = _get_or_create_direction(ifc, ax1[0]/s1, ax1[1]/s1, ax1[2]/s1)
+    d2 = _get_or_create_direction(ifc, ax2[0]/s2, ax2[1]/s2, ax2[2]/s2)
+    d3 = _get_or_create_direction(ifc, ax3[0]/s3, ax3[1]/s3, ax3[2]/s3)
 
-    # Translation, scaled to mm
-    tx = float(t[3])  * ts
-    ty = float(t[7])  * ts
-    tz = float(t[11]) * ts
+    # Translation, scaled and rounded to mm
+    tx = round(float(t[3])  * ts, 3)
+    ty = round(float(t[7])  * ts, 3)
+    tz = round(float(t[11]) * ts, 3)
     origin = ifc.createIfcCartesianPoint([tx, ty, tz])
 
-    # Use non-uniform variant to handle mirrors and non-uniform scale
+    # Round scales for cleaner output
+    s1 = round(s1, 6)
+    s2 = round(s2, 6)
+    s3 = round(s3, 6)
+
     return ifc.createIfcCartesianTransformationOperator3DnonUniform(
         d1,      # Axis1
         d2,      # Axis2
         origin,  # LocalOrigin
         s1,      # Scale
-        d3,      # Axis3
+        d3,      # Axis3 (explicit — never derived)
         s2,      # Scale2
         s3,      # Scale3
     )
@@ -529,12 +623,18 @@ def print_instance_stats():
     print(f"  Instance resolution: {_stats['found']}/{total} definitions found")
     if _stats["not_found"] > 0:
         print(f"  Warning: {_stats['not_found']} instances had no definition geometry")
+    unique_defs = len(_rep_map_cache)
+    unique_geom = len([v for v in _geometry_hash_cache.values() if v is not None])
+    if unique_defs > unique_geom:
+        print(f"  Geometry dedup: {unique_defs} definitions -> {unique_geom} unique geometries")
 
 
 def reset_caches():
     """Reset module-level caches (call at start of each export run)."""
     _mesh_data_cache.clear()
     _rep_map_cache.clear()
+    _geometry_hash_cache.clear()
     _identity_placement_cache.clear()
+    _direction_cache.clear()
     _stats["found"] = 0
     _stats["not_found"] = 0