37Curation Audit

Look at every picture. Then say what is wrong.

V1 assembled 40 artworks and 20 nature images and validated their rights. V1.1 looked at all 60 again — every source, every derivative, every record — and asked a harder question: does this work carry the thesis, and does it survive the way Milo actually shows a picture. It mostly does. Where it does not, it is written down here rather than quietly left in place.

IMAGES REVIEWED 60DERIVATIVES INSPECTED 240 + 20NOTHING DELETED

The verdicts37.1

Core · secondary · remove-replace.

CORE earns a hero. SECONDARY is usable but never large. REMOVE-REPLACE is demoted from the core library and superseded by a named candidate — not deleted from disk.

ART — 40 WORKS

22CoreCarries the thesis and survives all four recipes.

10SecondaryWorks at small size, in one recipe, or behind nothing.

8Remove / replaceDuplicated, unusable behind type, or collapses when treated.

NATURE — 20 IMAGES

9CoreEight from orbit, one from the ground.

4SecondaryReal, but the subject is not legible at a glance.

7Remove / replaceTravel brochure, archive filler, or a graphic baked into the pixels.

How each image was judged

Test	What was asked
Thesis	Which of the eleven qualities does the work actually carry — not which the theme field claims.
Recipe survival	Does the subject still read in all four recipes, or does national-park flatten it to a vermilion field and nous sink it to black.
Crop	Does it hold at a 16:9 band taken from its centre, which is how a hero or a card uses it.
Type	Can cream display type and carbon micro-type sit on the treated image and stay legible.
Duplication	Does another work already say this, by the same artist or of the same subject.

Nothing was judged from a filename or a metadata line. Two contact sheets of all 40 art thumbnails, then a comparison sheet per work holding the clean derivative, all four recipes and a type-over-image crop: 240 rendered images. For nature, all 20 thumbnails, then the derivatives — which changed two verdicts in both directions.

evidence/art-curation-v1_1.md · evidence/nature-curation-v1_1.md

What the audit taught the engine37.2

Four recipe behaviours that are system facts, not accidents.

These decide which images can be treated at all, and they are now selection criteria rather than surprises.

national-park quantises to four levels. A high-key or low-contrast source puts most of its luminance in the vermilion band and returns a near-solid vermilion field. Turner's Whalers, van Gogh's The Bedroom and Hiroshige's Sudden Shower all lose their subject this way.
Very dark sources go the other way. Rembrandt's Aristotle is nearly a black rectangle in both national-park and nous.
archive is close to a no-op on works already monochrome on paper — and turns colour paintings flat grey. It is doing its job in neither case.
A large plain ground defeats every recipe. The kouros's studio backdrop becomes an airbrush gradient; Audubon's cream sheet bleaches to blank paper.

This is also the argument for the sixth recipe. Some pictures are best served by almost no intervention at all.

Collection balance37.3

The hardest finding: one woman in forty.

Counted from assets/art/manifest.json, not from the curation document. A brand that says it champions artists cannot leave this unstated.

Artist gender1 of 40 by a woman (Cassatt) · 38 by named men · 1 anonymous. The named replacements take it to 6.

Geography2 non-Western works, both 19th-century Japanese prints. Nothing from China, Korea, South Asia, the Islamic world, Africa, Latin America or Indigenous North America — although all four institutions publish open-access work in every one of those areas.

Era31 of 40 made between 1830 and 1906 — 78% inside 76 years. Nothing medieval, nothing from the 18th century, nothing after 1906.

Medium29 paintings. 1 photograph, 1 sculpture, 8 works on paper — for a system whose image engine descends entirely from printmaking.

DuplicationNine artists supply 24 of the 40 works. Van Gogh 5, three of them trees in Provence in 1889.

Taking every replacement gives women 1 → 6, non-Western 2 → 4, Black and Native American artists 0 → 1, sculpture 1 → 3, works on paper 8 → 10. That is a start and not a fix: Africa, South Asia and Latin America are still empty afterwards, and are named here as V1.2 work rather than quietly closed.

A note on subject, not authorship

Seven of the 40 are American-sublime landscapes made by white artists for an expansionist audience, several of which include Indigenous people as scenery. Keeping them is defensible — they are the literal source of the national-park visual territory this system borrows. But the collection should not depict Indigenous people only through that lens while holding no work by an Indigenous artist.

What the validator does not catch37.4

A PASS is not the same as correct.

tools/validate_provenance.py reports 60 records, 0 problems — and it is right on its own terms. Every required field is present, every file exists, every derivative traces to an original. These are the defects it cannot see.

Defect	Records	Why it matters
Colour rendering is never declared	20 nature	A false-colour composite shown as a photograph is the same class of claim as an interpretation shown without its original — the thing this system exists to forbid. V1.1 adds colour_rendering, and writes unverified wherever the packaged caption does not say. One image the audit called false colour turned out to carry the word “true-color” in its own NASA caption.
Archive date recorded as capture date	11 nature	All eleven carried 2017-12-08, the day the item entered the NASA library. Nine real dates were recovered from the records’ own captions; two say nothing, and are left uncorrected rather than guessed.
Instrument attributed wrongly	2 nature	Both say “Landsat”. NASA's own captions say ASTER and MODIS. Carrying the agency's caption and then contradicting it is a defect.
Graphics baked into the source pixels	2 nature	A frame timestamp and a park boundary survive into the derivatives and read as though the engine drew them.
The two manifest twins disagree	1 nature	JSON says Denali, CSV says Glacier. One is wrong; the validator compares neither.
Slug names a nationality, not the artist	7 art	art-british-… is Turner. The id is the stable name of the thing, and the rule is one name per thing.
One artist, two creator strings	2 art	Breaks artist grouping and the About Artist action.
Below the documented 1200px	7 nature · 2 art	The package does not meet its own stated figure, because the chosen original was smaller.
No display title for art	4 art	Titles run 100–160 characters with archival apparatus. The caption sets the title in display serif and will overflow.

Full tables: evidence/art-curation-v1_1.md §7 · evidence/nature-curation-v1_1.md “Provenance weaknesses”.

What V1.1 actually did37.5

Recorded, not deleted.

An audit is not a purge. No image was removed from disk, no manifest record was dropped, and no replacement was fetched.

Every record carries a curation status	curation.status is core, secondary or replace, with the reason and, where one exists, the named candidate.
Factual errors corrected	The instrument names, the disagreeing location and the capture dates are facts, and facts get fixed rather than annotated.
Nothing fetched	Every replacement needs an object-level rights check and a confirmed accession number before it can be adopted. Naming a candidate is not adopting it.
Downstream use mapped	One demoted nature image appears 16 times across the site. Its replacement is chosen first, so no page is ever left without an image.

The point of writing it down this way is that the next person can disagree with a verdict without having to redo the looking.