Vanilla extract has 109 suppliers, and its odour field holds one word: vanilla

Ylang ylang grade I gets ten descriptors and 319 hours of substantivity data. Grades II and III, which have more suppliers between them, get two words: 'floral; ylang'. So I measured whether description depth tracks commercial importance. Overall it does — across the 6,913 records with an odour field, the median runs from 3 descriptors at zero suppliers to 8 at fifty-plus, Spearman ρ = 0.433. But 56 records break the line: twenty or more suppliers and only one or two descriptors. And 55.4% of those 56 have a descriptor that is simply their own name — vanilla extract says vanilla, peppermint oil says peppermint, garlic oil says garlic. In the well-described control group it is 22.0%. This field cannot describe the things that are themselves the definition.

About a 4 minute read.

Sweeping the FEMA duplicate groups last time, the ylang ylang group stopped me:

Suppliers	Record	Odour field	Substantivity
42	Ylang ylang flower oil I	floral, petal, jasmin, neroli, lily of the valley, spicy, balsamic, earthy, woody, winey	319 hours
30	Ylang ylang flower oil II	floral, ylang	— (empty)
45	Ylang ylang flower oil III	floral, ylang	— (empty)

Grade I gets ten descriptors; II and III get two each — and II and III together have more suppliers than I.

So I measured it: does this field's richness track how important a material is?

Overall: yes

The 6,913 records with an odour field, grouped by supplier count:

Suppliers	Records	Median descriptors	Has strength	Has substantivity
0	1,089	3	13.8%	12.9%
1–4	2,452	3	12.8%	11.9%
5–9	1,299	4	27.5%	26.6%
10–19	972	5	44.1%	42.3%
20–49	808	7	66.8%	66.1%
50–99	251	8	88.0%	88.0%
100+	42	8	85.7%	85.7%

Spearman ρ = 0.433. Descriptors climb from 3 to 8; the strength field goes from 13.8% filled to 88%.

The line is clear: things people buy are things somebody smelled.

But 56 records fall off it

Fifty-six records have twenty or more suppliers and only one or two descriptors. And they are not obscure:

Suppliers	Material	The entire odour field
109	Vanilla extract	"vanilla"
88	Peppermint oil	"peppermint"
71	Rose specialty	"rose"
67	Cassia oil	"cassia; cinnamon"
66	(±)-camphor	"camphoreous"
56	Jasmin absolute sambac	"floral; jasmin"
53	Garlic oil	"garlic"
47	Lemon oil Italy	"lemon"
45	Ginger root oil CO2 extract	"spicy; ginger"
45	Ylang ylang flower oil III	"floral; ylang"
41	Patchouli oil molecular distilled	"patchouli"
39	Orris rhizome concrete	"orris"
25	Roasted coffee bean extract	"coffee; coffee roasted coffee"

That last one finishes the argument. It writes its own name twice.

The pattern, measured

I defined self-reference as a descriptor overlapping the material's own name (stripping generic words like oil, absolute, extract before matching):

	Self-referential
≤2 descriptors and ≥20 suppliers (56 records)	55.4%
≥6 descriptors and ≥20 suppliers (781 records)	22.0%
	2.5×

More than half of the thinly described group is repeating its own name.

Why this happens

Because the vocabulary of odour description came from these materials.

The descriptor "rose" exists because roses do. The word "vanilla" followed the pod. When you set out to describe rose oil, the only accurate word available is the one it donated.

This is not the database being lazy. It is the boundary of a descriptive system. A vocabulary anchored in physical references cannot describe the references themselves.

And the evaluation field is no escape

I expected odor_descriptions — the signed evaluation entries — to fill in the detail. It does not:

	Median characters in evaluation field
Thin group (56 records)	46
Rich group (781 records)	364

The three actual entries:

Vanilla extract: at 100.00%. vanilla Peppermint oil: at 10.00% in dipropylene glycol. peppermint Garlic oil: at 0.01% in propylene glycol. intense garlic

Garlic gained the word "intense." Otherwise it is the same word with a concentration attached.

But the concentration is useful

All 56 record an evaluation concentration, and the spread is this:

Evaluation concentration	Records
0.01%	3
0.10%	2
1.00%	2
10.00%	10
100.00%	39

From 0.01% to 100% — a ten-thousand-fold range.

The three at the bottom are allyl isothiocyanate and two garlic oils; the top end is ylang ylang II and III, vanilla pod extract, white tea leaf extract.

This field failed to tell you what it smells like, but it told you how far to dilute it before finding out. In practice that is more often the number you actually need.

In practice

When a record has one or two descriptors, check the supplier count first. If it is well supplied, that is not missing data — the material is that word.
For these records, read the evaluation concentration, not the descriptors. 0.01% and 100% are entirely different ways of handling something.
Be careful with families that vary internally, like ylang ylang. Grade I has 319 hours of substantivity data and grades II and III have none — which does not mean they are short-lasting, only that nobody measured.

What this doesn't establish

Self-reference is judged by string overlap. I stripped generic words like oil/absolute/extract and required tokens longer than three characters, but it remains a crude approximation — "cassia; cinnamon" counts as a hit, and those two words relate in a way more complicated than the letters.
ρ = 0.433 is a correlation, not a cause. Well-supplied materials may simply have more complex smells.
Descriptor counts come from splitting on semicolons. Records may punctuate differently.
"The vocabulary is anchored in physical references" is my interpretation; the database does not explain where its descriptors came from.
The number 56 depends on my thresholds (≤2 descriptors, ≥20 suppliers). Different thresholds give different counts.
I have not smelled any of the materials above.