diff --git a/docs/CITYJSON_INTEROP.md b/docs/CITYJSON_INTEROP.md
index 107836e..03c4697 100644
--- a/docs/CITYJSON_INTEROP.md
+++ b/docs/CITYJSON_INTEROP.md
@@ -202,9 +202,10 @@ LIMIT 10;
 make test_full
 ```
 
-Four test files are gated on `require cityjson`
+Five test files are gated on `require cityjson`
 (`cityjson_interop`, `cityjson_hollow_solid`, `cityjson_multisolid`,
-`cityjson_delft_remote`), and one on `require spatial` (`spatial_coexist`).
+`cityjson_3dbag_attributes`, `cityjson_delft_remote`), and one on
+`require spatial` (`spatial_coexist`).
 Under `make test` / `make test_debug` they skip; `make test_full` does the setup
 that makes them execute, and treats a skip as a failure. It needs network access
 — the httpfs/spatial download on the first run, and the remote Delft fixture on
diff --git a/docs/TESTING.md b/docs/TESTING.md
index cc93117..1b75b44 100644
--- a/docs/TESTING.md
+++ b/docs/TESTING.md
@@ -523,7 +523,11 @@ FROM good WHERE n_parts = 1 AND b3_volume_lod22 > 0 AND b3_opp_grond > 0;
 ```
 
 The automated version of this check is
-[`test/sql/cityjson_delft_remote.test`](../test/sql/cityjson_delft_remote.test).
+[`test/sql/cityjson_delft_remote.test`](../test/sql/cityjson_delft_remote.test), which extends
+it to volume at LoD1.2 and LoD1.3, surface area against the five `b3_opp_*` attributes summed,
+`ST_3DZMin` against `b3_h_maaiveld` and the LoD1.2 extrusion height against `b3_h_dak_70p`.
+[`test/sql/cityjson_3dbag_attributes.test`](../test/sql/cityjson_3dbag_attributes.test) runs
+the same comparisons offline, over the frozen nine-building slice.
 
 ## 14 — Transforms: rigid motions and the cube law
 
@@ -1095,7 +1099,7 @@ The behaviours it demonstrates are covered automatically as follows:
 | §2 | `test/sql/cityjson_hollow_solid.test`, `test/sql/st_3d_hollow_solid.test`, `test/cpp/test_inner_shell.cpp` |
 | §3 | `test/sql/cityjson_multisolid.test`, `test/sql/st_3d_multisolid.test` |
 | §5, §6, §17 | `test/sql/geom_3d_accessors.test`, `geom_3d_measurements.test`, `geom_3d_serialize.test` |
-| §8–§13 | `test/sql/cityjson_delft_remote.test` (the 3DBAG oracle), `test/sql/st_3d_validation.test`, `test/sql/geom_3d_distance.test` |
+| §8–§13 | `test/sql/cityjson_delft_remote.test` and `test/sql/cityjson_3dbag_attributes.test` (the 3DBAG oracle, remote and offline), `test/sql/st_3d_validation.test`, `test/sql/geom_3d_distance.test` |
 | §14 | `test/sql/metamorphic_transforms.test` — **extended by this walkthrough** with the RotateX/RotateY cases from the quirks section |
 | §14 (kernel) | `test/cpp/test_measurements.cpp` — **added by this walkthrough**: far-from-origin precision |
 | §15 | `test/sql/st_transform.test`, `test/cpp/test_crs_transform.cpp` |
diff --git a/docs/TEST_COVERAGE.md b/docs/TEST_COVERAGE.md
index dd0ae1f..7f01787 100644
--- a/docs/TEST_COVERAGE.md
+++ b/docs/TEST_COVERAGE.md
@@ -12,7 +12,7 @@ Suite layout and the TDD workflow are in [AGENTS.md](../AGENTS.md); this file re
 | `test/cpp/` | `make test_cpp`, `make test_all`, `make test_full` | Kernel logic with no database: WKB parsing, model construction, payload round-trips, validation, triangulation, area/volume math |
 | `test/sql/` | `make test`, `make test_debug`, `make test_full` (release, gated tests included) | SQL surface: binding, null propagation, `TRY` semantics, result contracts, interop |
 
-Four tests gated on `require cityjson` skip unless a **locally built** `duckdb-cityjson` is
+Five tests gated on `require cityjson` skip unless a **locally built** `duckdb-cityjson` is
 staged for the sqllogic runner — the community-published extension emits an older column
 shape these tests no longer target. `make test_full` stages it (and fails the run on any
 skip); see [CITYJSON_INTEROP.md](./CITYJSON_INTEROP.md#running-the-gated-tests-under-sqllogic)
@@ -25,13 +25,40 @@ Three independent sources, each authoritative for a different class of claim.
 
 ### 3DBAG published attributes — real-geometry measurement
 
-`test/sql/cityjson_delft_remote.test` streams the 3DBAG Delft tile and compares
-`ST_3DVolume` against `b3_volume_lod22` and `ST_3DFootprintArea` against `b3_opp_grond`.
-This is the oracle of record for measurement on **real reconstructed geometry**, because
-SFCGAL rejects most real roofs for non-planarity while this extension measures them by
-triangulation.
+3DBAG publishes, per building, what its own reconstruction pipeline measured. Those figures
+are an independent toolchain's answer for the very geometry it shipped, which makes them the
+oracle of record for measurement on **real reconstructed geometry** — the case SFCGAL cannot
+serve, because it rejects most real roofs for non-planarity while this extension measures
+them by triangulation.
 
-Requires network access and the `cityjson` extension.
+| Attribute | Oracles |
+| --- | --- |
+| `b3_volume_lod12`, `b3_volume_lod13`, `b3_volume_lod22` | `ST_3DVolume`, at each LoD |
+| `b3_opp_grond` + `b3_opp_dak_plat` + `b3_opp_dak_schuin` + `b3_opp_buitenmuur` + `b3_opp_scheidingsmuur` | `ST_3DSurfaceArea` / `ST_3DArea` at LoD2.2 — every face of that solid carries exactly one of the five semantic roles, so they sum to its whole boundary |
+| `b3_opp_grond` | `ST_3DFootprintArea` — the ground surface is horizontal, so its 3D area and its XY projection are one number |
+| `b3_h_maaiveld` | `ST_3DZMin`, at every LoD |
+| `b3_h_dak_70p` | `ST_3DZMax` at LoD1.2 — the height the block was extruded to |
+| `b3_val3dity_lod22` | `ST_3DValidationReport` — one-way: what val3dity passes, this extension must pass |
+
+Two tests run those comparisons.
+
+- **`test/sql/cityjson_3dbag_attributes.test`** — offline, over the frozen nine-building slice
+  in `test/data/3dbag.city.jsonl`, at the tolerances that slice actually meets (0.1 % on
+  volume, 1 % on surface area, 2 mm on ZMin) plus string-exact per-building pins. Needs the
+  `cityjson` extension but **no network**, so it is the one that runs anywhere.
+- **`test/sql/cityjson_delft_remote.test`** — the same comparisons at tile scale, streamed
+  from the remote Delft tile. At that scale the thresholds are fractions ("≥95 % of buildings
+  within 2 %") rather than bounds, because a tile contains reconstructions that are simply
+  broken. Requires network access.
+
+The val3dity cross-check is a fraction, not a bound, and deliberately: the two disagree on a
+handful of tile solids that carry a single collapsed face. val3dity tolerates it; this
+extension reports the duplicated edge it creates as non-manifold and refuses to measure the
+solid, which is the documented no-silent-repair behaviour rather than a defect.
+
+Numeric pins in both files are formatted with `printf` and compared as text. sqllogic compares
+numeric results with a **1 % relative tolerance**, so a bare `ROUND(...)` pin admits drift an
+order of magnitude larger than the decimals it displays.
 
 ### PostGIS + SFCGAL — analytic measurement math
 
@@ -111,6 +138,7 @@ over the whole frozen 3DBAG corpus.
 | `ST_AsWKBMultiCube(separation)` | Volume conditioning when a collection's parts are spatially separated — the per-shell reference point of DESIGN_DOC §8.2. Total volume stays 16 at separations of 1e4–1e10, and the 1e6 case is oracled against SFCGAL as `fixture:multi_cube_far` |
 | `test/data/multisolid.city.json`, `compositesolid.city.json` | `MultiSolid` / `CompositeSolid` import per solid, with shell grouping recovered from the sidecar (`test/sql/cityjson_multisolid.test`). Import does **not** raise on either class |
 | `test/data/unit_cube.city.json` | End-to-end `cityjson` → `three_d` smoke test |
+| `test/data/3dbag.city.jsonl` | Nine real 3DBAG buildings at LoD1.2 / 1.3 / 2.2. The WKB the PostGIS oracle freezes as its `NL.IMBAG.Pand.*` rows comes from here, and `test/sql/cityjson_3dbag_attributes.test` measures the same nine against 3DBAG's published attributes offline |
 
 ## Open work
 
diff --git a/test/sql/cityjson_3dbag_attributes.test b/test/sql/cityjson_3dbag_attributes.test
new file mode 100644
index 0000000..343dfd4
--- /dev/null
+++ b/test/sql/cityjson_3dbag_attributes.test
@@ -0,0 +1,196 @@
+# name: test/sql/cityjson_3dbag_attributes.test
+# description: three_d measurements cross-checked against 3DBAG's own published attributes, on the frozen local tile slice
+# group: [sql]
+
+# ─────────────────────────────────────────────────────────────────────────────
+# 3DBAG publishes, per building, the figures its own reconstruction pipeline
+# measured: volume per LoD, the areas of ground / roof / wall surfaces, and the
+# ground elevation. Those are an oracle computed by an independent toolchain on
+# the same geometry, so they check `ST_3D*` measurement end to end.
+#
+# `test/data/3dbag.city.jsonl` is a frozen nine-building slice — the same nine
+# solids the PostGIS oracle freezes as its `NL.IMBAG.Pand.*` rows
+# (test/data/postgis_oracle/README.md). SFCGAL rejects eight of the nine for
+# non-planarity and measures neither their area nor their volume, so for those
+# this is the only oracle there is. It is also offline: unlike
+# cityjson_delft_remote.test it needs no network, so the tolerances are pinned
+# to what this frozen slice actually produces rather than left loose for drift.
+#
+# Which attribute checks which function:
+#
+#   b3_volume_lod12/13/22                         ST_3DVolume, per LoD
+#   b3_opp_grond + dak_plat + dak_schuin
+#     + buitenmuur + scheidingsmuur               ST_3DSurfaceArea at LoD2.2
+#   b3_opp_grond                                  ST_3DFootprintArea
+#   b3_h_maaiveld                                 ST_3DZMin
+#   b3_h_dak_70p                                  ST_3DZMax at LoD1.2
+#   b3_val3dity_lod22                             ST_3DValidationReport
+#
+# The area identity holds at LoD2.2 only: the five `b3_opp_*` attributes
+# describe the LoD2.2 model, and every face of that solid carries exactly one of
+# those semantic roles, so they sum to its whole boundary. The LoD1.2/1.3 blocks
+# extrude to a single roof height and their walls are a different surface.
+#
+# Requires the `cityjson` extension — a build of the sibling duckdb-cityjson
+# checkout, which emits the per-LoD geometry_lod<X> columns read below. See
+# docs/CITYJSON_INTEROP.md.
+# ─────────────────────────────────────────────────────────────────────────────
+
+require cityjson
+
+require three_d
+
+# One read yields every LoD present. Geometry lives on the `BuildingPart`
+# features and the attributes on the parent `Building`, so each part is joined
+# back to its parent.
+statement ok
+CREATE TABLE feats AS SELECT * FROM read_cityjsonseq('test/data/3dbag.city.jsonl');
+
+statement ok
+CREATE TABLE parts AS
+SELECT p.id,
+       ST_3DFromWKB(p.geometry_lod1_2, p.geometry_properties_lod1_2) AS s12,
+       ST_3DFromWKB(p.geometry_lod1_3, p.geometry_properties_lod1_3) AS s13,
+       ST_3DFromWKB(p.geometry_lod2_2, p.geometry_properties_lod2_2) AS s22,
+       b.b3_volume_lod12, b.b3_volume_lod13, b.b3_volume_lod22,
+       b.b3_opp_grond,
+       b.b3_opp_grond + b.b3_opp_dak_plat + b.b3_opp_dak_schuin
+         + b.b3_opp_buitenmuur + b.b3_opp_scheidingsmuur AS b3_opp_totaal,
+       b.b3_h_maaiveld, b.b3_h_dak_70p,
+       CAST(b.b3_val3dity_lod22 AS VARCHAR) AS b3_val3dity_lod22,
+       len(b.children) AS n_parts
+FROM feats p
+JOIN feats b ON b.id = p.parents[1]
+WHERE p.geometry_lod2_2 IS NOT NULL;
+
+# The slice is nine single-part buildings, each carrying all three LoDs, and
+# every geometry imports. Attributes describe the whole Building, so a
+# multi-part building would need its parts summed before any comparison below
+# held — this slice has none, and the assertion keeps it that way.
+query IIIII
+SELECT count(*), count(s12), count(s13), count(s22), max(n_parts) FROM parts;
+----
+9	9	9	9	1
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Volume, per LoD. Measured agreement on this slice is within 0.035 % at every
+# LoD; the 0.1 % bound leaves room for the vertex quantisation 3DBAG applies on
+# export (scale 0.001) without admitting a real regression.
+#
+# Each assertion lists the buildings that fail, so a failure names its rows.
+# ─────────────────────────────────────────────────────────────────────────────
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DVolume(s12) - b3_volume_lod12) > 0.001 * b3_volume_lod12
+ORDER BY id;
+----
+
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DVolume(s13) - b3_volume_lod13) > 0.001 * b3_volume_lod13
+ORDER BY id;
+----
+
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DVolume(s22) - b3_volume_lod22) > 0.001 * b3_volume_lod22
+ORDER BY id;
+----
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Surface area at LoD2.2 against the five semantic-surface areas summed. This is
+# the only oracle `ST_3DSurfaceArea` has on real reconstructed geometry.
+# Measured worst case on this slice is 0.55 %, on buildings where 3DBAG's
+# published roof area disagrees with the roof it shipped; 1 % is the bound.
+# ─────────────────────────────────────────────────────────────────────────────
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DSurfaceArea(s22) - b3_opp_totaal) > 0.01 * b3_opp_totaal
+ORDER BY id;
+----
+
+# Footprint area against the published ground area. The ground surface is
+# horizontal, so its 3D area and its XY projection are the same number, and
+# agreement is near-exact — within 0.005 m² measured, 0.02 m² asserted.
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DFootprintArea(s22) - b3_opp_grond) > 0.02
+ORDER BY id;
+----
+
+# The footprint does not move between LoDs: the same ground surface is carried
+# at every level, remeshed but not reshaped.
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DFootprintArea(s22) - ST_3DFootprintArea(s12)) > 0.01
+   OR abs(ST_3DFootprintArea(s22) - ST_3DFootprintArea(s13)) > 0.01
+ORDER BY id;
+----
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Heights. ZMin is the ground elevation 3DBAG publishes, to within the 1 mm
+# quantisation step of the file's transform. ZMax at LoD1.2 is the 70th
+# percentile roof elevation the block was extruded to — reconstruction snaps
+# that height, so it is checked at 5 cm, not at the quantisation step.
+# ─────────────────────────────────────────────────────────────────────────────
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DZMin(s22) - b3_h_maaiveld) > 0.002
+   OR abs(ST_3DZMin(s12) - b3_h_maaiveld) > 0.002
+   OR abs(ST_3DZMin(s13) - b3_h_maaiveld) > 0.002
+ORDER BY id;
+----
+
+query I
+SELECT id FROM parts
+WHERE abs(ST_3DZMax(s12) - b3_h_dak_70p) > 0.05
+ORDER BY id;
+----
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Validity. `b3_val3dity_lod22` is val3dity's error list for the LoD2.2 solid,
+# and `[]` means it found none. val3dity checks strictly more than this
+# extension does, so the implication runs one way only: what val3dity passes,
+# ST_3DValidationReport must also pass. All nine are clean under both.
+# ─────────────────────────────────────────────────────────────────────────────
+query I
+SELECT id FROM parts
+WHERE b3_val3dity_lod22 = '[]' AND NOT ST_3DValidationReport(s22).is_valid
+ORDER BY id;
+----
+
+query I
+SELECT count(*) FROM parts WHERE b3_val3dity_lod22 = '[]';
+----
+9
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Per-building pins. Face counts are structural — they come from the file, not
+# from a measurement. The measured figures are formatted to a fixed number of
+# decimals and compared as text: sqllogic compares numeric results with a 1 %
+# relative tolerance, which would wave through drift far larger than anything
+# the assertions above allow, so every pinned number here is a string.
+# ─────────────────────────────────────────────────────────────────────────────
+query TIITTTTTT
+SELECT id,
+       ST_3DNumFaces(s12), ST_3DNumFaces(s22),
+       printf('%.2f', ST_3DVolume(s22)),      printf('%.2f', b3_volume_lod22),
+       printf('%.2f', ST_3DSurfaceArea(s22)), printf('%.2f', ST_3DFootprintArea(s22)),
+       printf('%.3f', ST_3DZMin(s22)),        printf('%.3f', ST_3DZMax(s22))
+FROM parts ORDER BY id;
+----
+NL.IMBAG.Pand.0703100000027940-0	10	10	1399.54	1399.74	876.41	196.79	2.560	10.400
+NL.IMBAG.Pand.0703100000028278-0	15	40	3474.99	3475.19	1805.90	581.67	3.722	11.074
+NL.IMBAG.Pand.0703100000028817-0	6	6	1545.20	1545.00	1162.05	336.02	2.547	7.439
+NL.IMBAG.Pand.0703100000030681-0	6	6	550.97	551.01	470.15	88.74	-1.666	4.622
+NL.IMBAG.Pand.0703100000031557-0	6	6	46.82	46.82	79.10	15.21	2.835	5.921
+NL.IMBAG.Pand.0703100000031696-0	6	6	60.34	60.36	94.10	15.46	2.412	6.349
+NL.IMBAG.Pand.0703100000035296-0	8	8	37.25	37.26	70.32	14.86	4.329	6.861
+NL.IMBAG.Pand.0703100000035935-0	6	6	375.06	375.19	420.82	143.92	3.296	5.902
+NL.IMBAG.Pand.0703100000193621-0	8	12	1483.16	1483.55	1049.01	251.22	3.279	15.133
+
+statement ok
+DROP TABLE parts;
+
+statement ok
+DROP TABLE feats;
diff --git a/test/sql/cityjson_delft_remote.test b/test/sql/cityjson_delft_remote.test
index 8734bf5..bf040f5 100644
--- a/test/sql/cityjson_delft_remote.test
+++ b/test/sql/cityjson_delft_remote.test
@@ -5,12 +5,25 @@
 # ─────────────────────────────────────────────────────────────────────────────
 # This test streams real 3D building models — the 3DBAG "Delft" tile, CityJSON
 # 2.0 / EPSG:7415 — from a remote server and checks three_d's measurements
-# against 3DBAG's own published attributes (`b3_volume_lod22`, `b3_opp_grond`).
+# against the figures 3DBAG's own reconstruction pipeline published for them:
+#
+#   b3_volume_lod12/13/22                         ST_3DVolume, per LoD
+#   b3_opp_grond + dak_plat + dak_schuin
+#     + buitenmuur + scheidingsmuur               ST_3DSurfaceArea at LoD2.2
+#   b3_opp_grond                                  ST_3DFootprintArea
+#   b3_h_maaiveld                                 ST_3DZMin
+#   b3_h_dak_70p                                  ST_3DZMax at LoD1.2
+#   b3_val3dity_lod22                             ST_3DValidationReport
+#
+# `test/sql/cityjson_3dbag_attributes.test` runs the same comparisons offline,
+# against a frozen nine-building slice and at far tighter tolerances. This file
+# is the one that runs them at tile scale, where the thresholds have to be
+# fractions rather than bounds.
 #
 # It needs BOTH:
 #   1. the `cityjson` extension staged for the sqllogic runner — a build of the
 #      sibling duckdb-cityjson checkout, which emits the per-LoD
-#      `geometry_lod2_2` / `geometry_properties_lod2_2` columns aliased below
+#      `geometry_lod<X>` / `geometry_properties_lod<X>` columns read below
 #      (the published community build's flat `geometry` will not bind), and
 #   2. network access to
 #      https://cityjson.open3d.city/cityjsonseq/delft.city.jsonl
@@ -36,23 +49,54 @@ statement ok
 CREATE TABLE feats AS
 SELECT id, parents, children,
        geometry_lod2_2 AS geometry, geometry_properties_lod2_2 AS geometry_properties,
-       b3_volume_lod22, b3_opp_grond
-FROM read_cityjsonseq(
-    'https://cityjson.open3d.city/cityjsonseq/delft.city.jsonl',
-    lod => '2.2')
+       geometry_lod1_2, geometry_properties_lod1_2,
+       geometry_lod1_3, geometry_properties_lod1_3,
+       b3_volume_lod12, b3_volume_lod13, b3_volume_lod22,
+       b3_opp_grond,
+       b3_opp_grond + b3_opp_dak_plat + b3_opp_dak_schuin
+         + b3_opp_buitenmuur + b3_opp_scheidingsmuur AS b3_opp_totaal,
+       b3_h_maaiveld, b3_h_dak_70p, b3_val3dity_lod22
+FROM read_cityjsonseq('https://cityjson.open3d.city/cityjsonseq/delft.city.jsonl')
 LIMIT 400;
 
 statement ok
 CREATE TABLE parts AS
 SELECT p.id,
        ST_3DTryFromWKB(p.geometry, p.geometry_properties) AS solid,
+       ST_3DTryFromWKB(p.geometry_lod1_2, p.geometry_properties_lod1_2) AS solid12,
+       ST_3DTryFromWKB(p.geometry_lod1_3, p.geometry_properties_lod1_3) AS solid13,
+       b.b3_volume_lod12,
+       b.b3_volume_lod13,
        b.b3_volume_lod22,
        b.b3_opp_grond,
+       b.b3_opp_totaal,
+       b.b3_h_maaiveld,
+       b.b3_h_dak_70p,
+       b.b3_val3dity_lod22,
        len(b.children) AS n_parts
 FROM feats p
 JOIN feats b ON b.id = p.parents[1]
 WHERE p.geometry IS NOT NULL;
 
+# The measurement assertions below run over the LoD1.2, LoD1.3 and LoD2.2 solids
+# of the same building, so restrict to the parts that carry, and import, all
+# three. ST_3DVolume and ST_3DSurfaceArea raise rather than repair, so the
+# volume and area comparisons are further restricted to the solids this
+# extension reports as valid.
+statement ok
+CREATE TABLE triples AS
+SELECT *, ST_3DValidationReport(solid).is_valid AS valid22
+FROM parts
+WHERE n_parts = 1
+  AND solid IS NOT NULL AND solid12 IS NOT NULL AND solid13 IS NOT NULL;
+
+statement ok
+CREATE TABLE valid_triples AS
+SELECT * FROM triples
+WHERE valid22
+  AND ST_3DValidationReport(solid12).is_valid
+  AND ST_3DValidationReport(solid13).is_valid;
+
 # Sanity: we actually pulled a non-trivial number of buildings.
 query I
 SELECT count(*) > 100 FROM parts;
@@ -108,17 +152,123 @@ SELECT bool_and((ST_3DZMax(solid) - ST_3DZMin(solid)) > 0) FROM parts;
 ----
 true
 
+# Every part carries all three LoDs, and all three import.
+query I
+SELECT count(*) > 100 FROM triples;
+----
+true
+
+# GROUND TRUTH — volume at LoD1.2 and LoD1.3. The block models are a separate
+# reconstruction from the LoD2.2 roof, so they exercise ST_3DVolume on geometry
+# the LoD2.2 assertion above never sees. Measured agreement is within 0.08 %.
+query I
+SELECT (count(*) FILTER (
+          WHERE abs(ST_3DVolume(solid12) - b3_volume_lod12)
+                <= 0.02 * b3_volume_lod12 + 0.1))
+       >= 0.95 * count(*)
+FROM valid_triples
+WHERE b3_volume_lod12 > 0;
+----
+true
+
+query I
+SELECT (count(*) FILTER (
+          WHERE abs(ST_3DVolume(solid13) - b3_volume_lod13)
+                <= 0.02 * b3_volume_lod13 + 0.1))
+       >= 0.95 * count(*)
+FROM valid_triples
+WHERE b3_volume_lod13 > 0;
+----
+true
+
+# GROUND TRUTH — surface area. Every face of the LoD2.2 solid carries exactly one
+# semantic role, so 3DBAG's five published surface areas — ground, flat roof,
+# sloped roof, exterior wall, party wall — sum to its whole boundary. This is
+# the only check ST_3DSurfaceArea has against an independent measurement of real
+# reconstructed geometry: SFCGAL rejects these roofs for non-planarity. Measured
+# median disagreement is 0.007 %, with ~99 % of buildings inside 2 %.
+query I
+SELECT (count(*) FILTER (
+          WHERE abs(ST_3DSurfaceArea(solid) - b3_opp_totaal)
+                <= 0.02 * b3_opp_totaal + 0.1))
+       >= 0.95 * count(*)
+FROM valid_triples
+WHERE b3_opp_totaal > 0;
+----
+true
+
+# GROUND TRUTH — ground elevation. b3_h_maaiveld is the height 3DBAG places the
+# ground surface at, and every LoD is built down to it, so ST_3DZMin must return
+# it — to within the 1 mm quantisation step of the tile's transform. On the
+# extruded blocks that holds for every building, not a fraction of them.
+query I
+SELECT bool_and(abs(ST_3DZMin(solid12) - b3_h_maaiveld) <= 0.01
+            AND abs(ST_3DZMin(solid13) - b3_h_maaiveld) <= 0.01)
+FROM triples;
+----
+true
+
+# At LoD2.2 it holds over the solids that reconstruct validly. It is stated over
+# those rather than over all of them because a broken LoD2.2 roof can drag a
+# vertex metres below the ground surface — the tile carries one such building,
+# and this extension reports it as invalid rather than measuring it.
+query I
+SELECT bool_and(abs(ST_3DZMin(solid) - b3_h_maaiveld) <= 0.01)
+FROM valid_triples;
+----
+true
+
+# GROUND TRUTH — extrusion height. The LoD1.2 block is extruded to the 70th
+# percentile roof elevation, so ST_3DZMax on it recovers b3_h_dak_70p.
+# Reconstruction snaps that height, so this is a decimetre-scale agreement over
+# most buildings rather than an exact one — measured 95th percentile 9 cm.
+query I
+SELECT (count(*) FILTER (WHERE abs(ST_3DZMax(solid12) - b3_h_dak_70p) <= 0.15))
+       >= 0.90 * count(*)
+FROM triples;
+----
+true
+
+# The footprint does not move between LoDs: the same ground surface is carried
+# at every level, remeshed but not reshaped.
+query I
+SELECT (count(*) FILTER (
+          WHERE abs(ST_3DFootprintArea(solid) - ST_3DFootprintArea(solid12))
+                <= 0.01 * ST_3DFootprintArea(solid12)))
+       = count(*)
+FROM triples
+WHERE ST_3DFootprintArea(solid12) > 0;
+----
+true
+
+# CROSS-CHECK — validity. b3_val3dity_lod22 is val3dity's error list for the
+# LoD2.2 solid; `[]` means it found none. val3dity checks strictly more than this
+# extension does, so the implication runs one way: what val3dity passes,
+# ST_3DValidationReport should pass too. It is a fraction rather than a bound
+# because the two disagree on a handful of solids that carry a single collapsed
+# face — val3dity tolerates it, this extension reports the duplicated edge it
+# creates as non-manifold and refuses to measure the solid. Measured 98.5 %.
+query I
+SELECT (count(*) FILTER (WHERE valid22)) >= 0.95 * count(*)
+FROM triples
+WHERE b3_val3dity_lod22 = '[]';
+----
+true
+
 # ─────────────────────────────────────────────────────────────────────────────
 # Deterministic per-building pins (selected by id, independent of row order).
 # NL.IMBAG.Pand.0503100000012869-0 is a small flat-roof house whose 3DBAG
 # LoD2.2 solid has 6 faces; published b3_volume_lod22 ≈ 19.52, b3_opp_grond ≈ 7.21.
 # ─────────────────────────────────────────────────────────────────────────────
-query IIIII
+# The measured figures are formatted and compared as text: sqllogic compares
+# numeric results with a 1 % relative tolerance, so a bare ROUND(...) pin would
+# wave through drift an order of magnitude larger than the rounding it shows.
+query IITTT
 SELECT ST_3DNumFaces(solid),
        ST_3DIsClosed(solid),
-       ROUND(ST_3DVolume(solid), 1),
-       ROUND(ST_3DFootprintArea(solid), 1),
-       ROUND(ST_3DZMax(solid) - ST_3DZMin(solid), 2)
+       printf('%.1f', ST_3DVolume(solid)),
+       printf('%.1f', ST_3DFootprintArea(solid)),
+       printf('%.2f', ST_3DZMax(solid) - ST_3DZMin(solid))
 FROM parts
 WHERE id = 'NL.IMBAG.Pand.0503100000012869-0';
 ----
@@ -134,6 +284,12 @@ SELECT ROUND(ST_3DDistance(
 ----
 true
 
+statement ok
+DROP TABLE valid_triples;
+
+statement ok
+DROP TABLE triples;
+
 statement ok
 DROP TABLE parts;