# The measured dimensions of catalogued paintings

720,147 rows: every catalogued painting in four open sources
whose picture-surface dimensions could be read out of the catalogue's own prose, from 736,698 candidate records across four sources.

**The unit is a catalogue record, not a painting**, and the difference is not pedantry: a
percentage computed here is a percentage of *catalogued* works, which is a different
quantity from a percentage of paintings and cannot stand in for it.

## Read this before you compute anything

**Three of the four sources are single museums. The fourth is Wikidata, and it is
97.8 per cent of the file.**

| `catalogue` | source | rows |
|---|---|---|
| `aic` | Art Institute of Chicago | 3,526 |
| `met` | The Metropolitan Museum of Art | 8,324 |
| `nga` | National Gallery of Art, Washington | 4,178 |
| `wd` | Wikidata (crowd-sourced) | 704,119 |

Any statistic taken over the whole file is a statistic about Wikidata with a rounding
error of three museums attached. **Group by `catalogue` and report separately.** The
instrument this came from does exactly that and never pools them.

## The limitation that matters most

**These rows carry no identifier.** No accession number, no object id, no Q-id. The
committed record set never retained one, so a row cannot be joined back to its catalogue
record and **you cannot check any single row against its source.** You can check the
aggregate; you cannot check the instance. Rebuilding with identifiers is possible, since
all four sources are CC0 and all four expose one, and costs about two hours of refetching.
It is not done here, and this paragraph exists so that nobody discovers it the hard way.

## What the parse decides

A dimension string like *"39 3/8 x 32 in. (100 x 81.3 cm), framed: 48 x 40 in."* has to be
resolved into one pair that is the picture surface. Records were rejected where that could
not be done honestly:

- **aic** (3,900 candidates): three-dimensional 87, multi-part-work 66, only-frame-or-mount 14, approximate-or-not-rectangular 22, no-string 128, ambiguous-multiple 16, no-pair 41
- **met** (9,005 candidates): approximate-or-not-rectangular 248, only-frame-or-mount 67, unit-mismatch 28, no-pair 119, ambiguous-multiple 81, multi-part-work 64, three-dimensional 60, no-string 14
- **nga** (4,446 candidates): ambiguous-multiple 59, approximate-or-not-rectangular 63, three-dimensional 83, only-frame-or-mount 7, multi-part-work 41, no-string 15
- **wd** (719,347 rows): duplicate 14,723, implausible-ratio 367, nonpositive 138

## Units

`unit` is null on every Wikidata row, because the query returns truthy values without the
unit qualifier. **Absolute sizes are therefore not comparable across the file.** `ratio`
is, because a ratio has no units, which is why the instrument is built on it.

## Files

- `paintings-aic.csv`, `paintings-met.csv`, `paintings-nga.csv`,
  `paintings-wd-1.csv`, `paintings-wd-2.csv`, `paintings-wd-3.csv`: RFC 4180 CSV,
  one file per catalogue, each with its own header line. UTF-8, LF, and an empty cell
  means null.
- **The split is deliberate.** These arrive as separate files because they must not be
  pooled, and because Wikidata alone is too large for one. Concatenating them is the one
  thing this dataset asks you not to do without saying so in your method.
- CSV rather than JSONL because every field is a flat scalar: it opens in pandas, R or a
  spreadsheet with no parsing, and `schema.json` is still the row contract.
- `schema.json`, `sources.json`, `manifest.json`, `validate.mjs`.

## Citation

> Artificial Wasteland (2026). *The measured dimensions of catalogued paintings.*
> https://artwaste.land/data/canvas-proportions/ · dimensions from the Art Institute of
> Chicago, The Metropolitan Museum of Art, the National Gallery of Art (Washington) and
> Wikidata, all CC0; dimension parse by the Artificial Wasteland, CC0 1.0.