This is not a list of logos. For each dataset here is what we measured ourselves: how many features were read, how many layers, whether the circle closed, what shift the coordinate came out with — and when it was measured.
28 datasetsread ourselves, each with a measurement date
26 publisherseach has its own manner, and rules are fixed to it
7 circles closedrecorded and read our own result back; of them 1 — with losses named in advance
24 trapsrecorded along the way: what these datasets break on — that is the experience itself
Numbers of the summary are about the registry file, not about your data. Measurements in it run from 2026-09-10 to 2026-09-23. The publisher may have re-released a dataset after the measurement date, and then our knowledge went stale silently — that is why the date stands next to every number below, not once per page.
How to read the outcome
The circle closed — we read, wrote, read our own result back and checked: datasets, fields, types, feature count, coordinates. No discrepancies.
The circle closed with named losses — the same, but the format does not do something, and we said so as a number before writing, then checked the promise against the measurement.
Read, circle not closed — the numbers on input were taken, no read-back check was done. This is “we read it”, not “it will work for you”.
We refused — we hit our own limit: memory or feature count. The reason is named in the measurement line, and it is ours to fix.
Nothing we have can read this — a refusal on the side of a third-party tool or the publisher. The reason is also named, but it is not ours to fix.
Circle closed: wrote and read our own result — 6 datasets
City of Prince George
CA-BC · arcgis.com · prince-george-standards
CoPG_Civil3D_Standards.zip (DWG, 565 files in the archive)
Arrived under the name opendatapg_cityofpg. The account name on the portal is not the publisher.
2026-09-22 · circle closed · through our server — the same path any person would use
6 786 features at input
6 786 features at output
245 layers
0.0000 m — coordinate shift
565 files in the archive
412 source drawings
153 non-drawings skipped
409 files merged into one dataset
3 files not read
the largest intake ever: 136 times more files than the previous record, and the circle closed down to the feature
2 traps of this dataset
The archive holds images, fonts and Excel tables mixed in with drawings. Refusing on the first non-drawing threw away 411 already-read ones.
.xlsx is NOT distinguishable from an archive of drawings by its first bytes — it's a zip, and the “is the file recognized” filter let it through: a job of 414 drawings failed on the 291st.
ParkerWater
US · arcgis.com · parker-water
ParkerWater_developer_dwg.zip (DWG)
2026-09-22 · circle closed · through our server — the same path any person would use
151 features at input
151 features at output
8 layers
0.0000 m — coordinate shift
sam.cragun
US · arcgis.com · newman-topo
Newman_Topo.zip (DWG)
2026-09-22 · circle closed · through our server — the same path any person would use
286 features at input
286 features at output
2 layers
0.0000 m — coordinate shift
EsriEgitim
TR · arcgis.com · heybeliada-cad-bim
Heybeliada_CAD_BIM.zip (DWG)
2026-09-22 · circle closed · through our server — the same path any person would use
52 487 features at input
52 487 features at output
38 layers
0.0000 m — coordinate shift
sam.cragun
US · arcgis.com · glm2019-contours
GLM2019_2m_contours_dxf.zip (DXF)
2026-09-22 · circle closed · through our server — the same path any person would use
2026-09-19 · circle closed · in GPKG · writing and read-back check directly
7 152 features at input
7 152 features at output
0.0000 m — coordinate shift
2026-09-19 · circle closed with losses named · in MapInfo File · writing and read-back check directly
3 355 features lost height
the loss was predicted before writing and matched by the number: promised 3355, got 3355
2026-09-10 · read, circle not closed · reading only
7 152 features
9 layers
3 traps of this dataset
The coordinate system is written inside the drawing as an ordinary text label: the Disclaimer layer holds a TEXT reading «data is setup to project to Nad83, Zone 10». Text needs not just carrying over, but reading.
Long multi-line TEXT is lost at the libredwg → GDAL seam: through DXF from libredwg 7149 features show, through DXF from ODA — 7152. Worked around with our own splicing (phase0/fixdxf.py); no need to buy ODA for this.
3355 of 7152 features have height — our main example of loss when writing to MapInfo and Shapefile.
Circle closed with losses named before writing — 1 dataset
Dublin
IE · dublin-mapinfo
ACA.TAB (MapInfo File)
2026-09-18 · circle closed with losses named · writing and read-back check directly
0.0100 m — coordinate shift
MERGED_shapes, 6555 polygons — shift 0.0100; MERGED_text — 0.0081; ACA and MERGED_lines with a recorded CRS — 0.0000
2 traps of this dataset
A format's “file” is not one file: .tab plus .dat, .map and .id. Bring just the one named file, and you get GDAL's refusal “not recognized as being in a supported file format” — a message about an UNSUPPORTED FORMAT exactly where the format reads perfectly well, and what was brought was an incomplete set.
MapInfo's coordinate sits in a grid tied to the table's own bounds. If the CRS isn't recorded, a 0.0100 shift appears even where it wasn't present with a recorded CRS: the CRS is not a label on the data, it's part of its geometry.
Read — circle not closed — 13 datasets
rockcogis
US · arcgis.com · rockco-soils
Soils CAD Format(DWG).DWG (DWG, 19.9 MB)
2026-09-18 · read, circle not closed · reading only
10 748 features
1 layers
chousa
JP · arcgis.com · chousa-senkei
線形.dwg (DWG, 486.7 KB)
2026-09-18 · read, circle not closed · reading only
827 features
layer names partly recoverable
1 trap of this dataset
The file name INSIDE the zip is written in Shift-JIS with no UTF-8 flag, and zipfile parses it as cp437. Breaks before any CAD step; fixed by re-encoding back to cp437 and parsing as cp932.
Hlavné mesto Bratislava
SK · arcgis.com · bratislava-funkcne-plochy
funkcne_plochy_UPN.dwg (DWG, 7.1 MB)
Arrived under the name data.bratislava_2. The account name on the portal is not the publisher.
2026-09-18 · read, circle not closed · reading only
8 977 features
1 layers
Miejska Pracownia Urbanistyczna w Łodzi
PL · arcgis.com · lodz-plans
Plany_przeznaczenia_cale_miasto.dxf (DXF, 14.3 MB, hash taken 2026-09-23 from /data/blind6/raw on the guest: name and size in bytes matched the record. The hash was not recorded at measurement time — this is identification after the fact, not that day's own record)
Arrived under the name m.brodowicz_MPULODZ1. The account name on the portal is not the publisher.
Recognized by layers: Granica miasta Łodzi, Granice planów obowiązujących, Granice przystąpień, Udostępniane granice terenów w planach miejscowych - dla planów uchwalanych od 2017
2026-09-17 · read, circle not closed · reading only
layer names recovered after fixing the encoding, except where a byte was lost
2 traps of this dataset
$DWGCODEPAGE is declared ANSI_1250, but the file is written in UTF-8. Layer names arrive looking like «Granice planow obowiazujacych» with garbage instead of diacritics. GDAL prints the warning ONCE and goes quiet. Fixed before classification (ingest.unmojibake), and the fix is checkable: re-encoding back gives valid UTF-8 only when UTF-8 was really there.
Part of the names cannot be fixed in principle: «Ł» in UTF-8 is 0xC5 0x81, and 0x81 is undefined in cp1250 and is lost for good in GDAL. Even such cases can be taken, but only by parsing the DXF with our own reader.
Township of Langley
CA-BC · arcgis.com · langley-pipes
Drn_Pipe_CAD.dwg (DWG); San_Pipe_CAD.dwg (DWG)
Arrived under the name OpenDataFME. The account name on the portal is not the publisher.
2026-09-17 · read, circle not closed · reading only
ten out of ten markings correct — the best result ever, and exactly why it wasn't counted as blind
1 trap of this dataset
A portal account name is not the publisher. Both files sit under the OpenDataFME account, while the publisher is the same Township of Langley; only an aside inside the accompanying .txt gave it away. Hence also the best run numbers: the rules had already been fixed for this style, so it wasn't blind after all.
Forsyth County, GA
US-GA · arcgis.com · forsyth-template
FORSYTHTEMPLATE 12-13-16.dwg (DWG, 19.7 KB, hash taken 2026-09-23 from /data/blind6/raw on the guest: name and size in bytes matched the record. The hash was not recorded at measurement time — this is identification after the fact, not that day's own record)
Arrived under the name FCPublishData. The account name on the portal is not the publisher.
2026-09-17 · read, circle not closed · reading only
1 features
40 layers declared
1 trap of this dataset
This is a TEMPLATE, not data: 40 layers declared, one feature. A dataset where many layers are declared but there are no features is a reason to say “this is a template”, not “we found nothing”.
Curicó
CL · arcgis.com · curico-plan-regulador
Plan Regulador curico.dwg (DWG, 1.1 MB, hash taken 2026-09-23 from /data/blind6/raw on the guest: name and size in bytes matched the record. The hash was not recorded at measurement time — this is identification after the fact, not that day's own record)
Arrived under the name NicolasPereiraL. The account name on the portal is not the publisher.
2026-09-17 · read, circle not closed · reading only
13 628 features
City of Greater Sudbury
CA-ON · sudbury-cityview
2009-CityView-3D-Vectors-all.dwg (DWG)
2026-09-15 · read, circle not closed · reading only
27 241 features
via GeoJSON; via JSON — a syntactically invalid document
1 trap of this dataset
dwgread -O JSON cuts the document off at byte 451,763,361 of 452,385,643: zero return code, empty stderr, the file is created. dwgread -O GeoJSON on the same file returns all the features. One output format's failure is not the tool's failure, and parsing the structure must not be a condition for reading the data.
Leon County, FL
US-FL · leon-lcstseg
LCSTSEG.dxf (DXF, AC1027, 55.5 MB)
Recognized by layers: Lcstseg
2026-09-14 · read, circle not closed · the file was taken into the test corpus
Delaware County, OH
US-OH · delaware-street-centerlines
StreetCenterlines.dxf (DXF, AC1027, 15.2 MB)
Recognized by layers: StreetCenterlines
2026-09-14 · read, circle not closed · the file was taken into the test corpus
1 trap of this dataset
The coordinate system sits in a separate .prj: NAD_1983_StatePlane_Ohio_North_FIPS_3401_Feet. The only file in the corpus where the CRS is stated explicitly and alongside.
2026-09-14 · read, circle not closed · the file was taken into the test corpus
City of Ottawa
CA-ON · ottawa-old-east-south
OLD OTTAWA EAST.dxf (DXF, AC1032, 26.5 MB); OLD OTTAWA SOUTH.dxf (DXF, AC1032, 25.6 MB)
Recognized by layers: OLD OTTAWA EAST, OLD OTTAWA SOUTH
2026-09-10 · read, circle not closed · the file was taken into the test corpus
Leon County / City of Tallahassee
US-FL · arcgis.com · leon-map-tiles
MapTiles.DWG (DWG, AC1015, 274.2 KB, hash matched corpus/MANIFEST.md and the file /data/raw/MapTiles.DWG on the guest, checked 2026-09-23); NE_Quadrant.zip (DWG, 199 tiles, unpacked 1,656,952,151 B)
Recognized by layers: MapTiles, MapTilesAnno
2026-09-10 · read, circle not closed · the file was taken into the test corpus
measured DWG→DXF inflation ×2.15 on 199 tiles, peak per pass 43 MB
2 traps of this dataset
A license with the reservation “for reference purposes only” — verbatim in corpus/MANIFEST.md.
The coordinate system is not set at all.
We refused, and here is why — 3 datasets
WDC-GIS
US · arcgis.com · wdc-assets
WDCAssets.dwg (DWG, AC1027, 20.6 MB)
2026-09-22 · we refused · through our server — the same path any person would use
233 995 features
Why: the file didn't fit in the stand's memory — killed in the converter, code −9
1 trap of this dataset
A drawing's weight comes from the number of features, not megabytes: 20.6 MB and 233,995 features. Memory runs out in TWO places — in dwgread while converting DWG to DXF (someone else's code, nothing to measure it with in advance) and in our own model (there a limit by count is in effect). Whether it “will fit” cannot be predicted at all before starting; a limit by size remains only an upper cutoff.
Hoogheemraadschap van Delfland
NL · arcgis.com · delfland-legger
Legger_Delfland_dxf.zip (DXF, 8 files, the largest 115,613,856 B)
Arrived under the name hhdelfland, leggerdelfland@hhdelfland.nl. The account name on the portal is not the publisher.
2026-09-22 · we refused · through our server — the same path any person would use
182 219 features
Why: the drawing has 182,219 features, and the stand takes on no more than 161,480 — refused in advance and by the number
2026-09-17 · read, circle not closed · reading only
182 219 features
all 182,219 on one layer, Ondersteunende kunstwerken
Porirua City Council
NZ · arcgis.com · porirua-contour
Contour_5m_2023.dwg (DWG, 432 MB)
2026-09-18 · we refused · reading only
Why: killed by memory in the converter, code 137
1 trap of this dataset
dwg2dxf grows to 2.5 GB and is killed by the container limit: code 137, not a single line in the log. This is also where our own silent failure surfaced — the runner was losing the return code in the pipe to tail.
Nothing we have can read this — 5 datasets
City of Brampton
CA-ON · brampton-roads
COB24_ROADS.dwg (DWG, AC1027, 50.2 MB)
2026-09-23 · nothing we have can read this · reading only
Why: intake was killed in the container by memory (code −9) after 272 s at a 2.2 GB ceiling. On the stand the job ceiling is 1.2 GB (lowered from 2.5 GB on 2026-09-21 for isolation and two workers) — there, this file does not read at all right now. On September 20 the same file completed a full circle through the web, but at a 2.5 GB ceiling, and intake took 1408 s
2026-09-20 · circle closed · writing and read-back check directly
353 636 features at input
353 636 features at output
200 000 features checked
2 traps of this dataset
The comparison takes 200,000 features out of 353,636 — this is OUR OWN limit per dataset, not a format refusal. About the rest, it is only stated that they are in place and counted.
Intake on this guest takes about 1408 s. This is the cost of re-reading, which is why the run's second stage is marked with the “already read” key.
cctgis
US · arcgis.com · warrenton-cad
Warrenton_CAD.zip (DWG)
2026-09-22 · nothing we have can read this · through our server — the same path any person would use
Why: both converter paths reported success while handing over emptiness
1 trap of this dataset
The same failure as Wellington's: dwg2dxf returns 0 and writes a 0 B DXF out of 128 MB, dwgread -O GeoJSON creates no output, both links report success.
Ventanilla
PE · ventanilla-plan
ventanilla.dxf (DXF)
2026-09-18 · nothing we have can read this · reading only
Why: broke off mid-traversal at the 77,274th feature out of 235,495 declared
1 trap of this dataset
A read probe is only as honest as it is deep: GetFeatureCount passes and says 235,495, and GetNextFeature fails on the 77,274th (ogrdxflayer.cpp, 2175). A break mid-traversal is a separate event: what was written must be DISCARDED and read again from zero by a fallback path, or a fragment rides out passed off as the whole file.
Tartu linn
EE · arcgis.com · tartu-maakasutus
maakasutus_yp_2040.dgn (DGN, 8, 1.5 MB)
Arrived under the name Tartu_admin. The account name on the portal is not the publisher.
2026-09-18 · nothing we have can read this · reading only
Why: DGN v8: no reader in the image
1 trap of this dataset
The first bytes are d0 cf 11 e0 a1 b1 1a e1 — an OLE2 container, i.e. DGN version 8. GDAL's DGNv8 driver only builds with Teigha, which our image doesn't have. The format is recognized but not read, and those are different statements.
Wellington City Council
NZ · arcgis.com · wellington-1m
Wellington1m.DWG (DWG, AC1027, 392.2 MB, hash taken 2026-09-23 from /data/blind6/raw on the guest: name and size in bytes matched the record. The hash was not recorded at measurement time — this is identification after the fact, not that day's own record)
2026-09-17 · nothing we have can read this · reading only
Why: both converter paths handed over emptiness and both reported success
1 trap of this dataset
Both converter paths return EMPTY and both say SUCCESS: dwg2dxf writes 8,845 B (a break at the start of TABLES) and returns 0; dwgread -O GeoJSON creates a zero-length file and prints SUCCESS. The refusal is named on the converter's side — “handed over a scrap and called it success” — not on GDAL's.
The registry lives in the repository as a file and is diffed. There is no data — yours or anyone else's — in it: only a fingerprint by layer names, traps in words and measurement numbers.