One more correction to last time's table: the assay field's coverage is not 64.5%, it is 10.5%

About a 4 minute read.

Article 100 listed coverage for thirty fields and caught four placeholder fields. Then I ran the same sweep across every field, and found an error in my own table.

The assay field

I reported 64.5% (27,221 records with a value).

But 22,771 of those carry the identical string "95.00 to 100.00." That is 83.7% of them.

assay value	Records
95.00 to 100.00	22,771
99.00 to 100.00	1,084
98.00 to 100.00	1,005
97.00 to 100.00	647
96.00 to 100.00	331

Strip that default and 4,450 records remain — 10.5% of the database.

I ranked it sixth in article 100. It belongs around eighteenth.

And the default follows a rule

It tracks almost perfectly with whether anyone sells the material:

	Records with assay	Share that are the default
Zero suppliers	13,382	97.5%
Marked natural	12,998	88.0%
Marked synthetic	633	56.6%
10+ suppliers	2,091	26.5%

Materials people sell have a real purity specification three times in four. Materials nobody sells carry boilerplate thirty-nine times in forty.

(88% boilerplate among naturals makes sense too — a natural extract has no "purity"; the field was never meaningful for it.)

So the right way to read the assay field is to check the supplier count first. For well-supplied materials it is a specification; for zero-supplier records it is filler.

The second kind of empty: the field holds a button label

I missed this entirely in article 100.

Field	With a value	Most common value	Share
chembl	5,831	"View"	100.0%
chebi	4,638	"View"	100.0%
fda_unii	30,358	"Search"	65.0%
hmdb	23,589	"Search"	68.8%

In chembl and chebi, every single record says "View."

Whatever scraped these pages captured the link's anchor text instead of the identifier it pointed to. "View" is a button, not an identifier.

fda_unii and hmdb are mixed — about a third are real identifiers and two thirds are "Search."

The control: kegg, foodb, nikkaji, beilstein and mdl all have a most-common value appearing 24 times or fewer. Those are real identifiers.

The third kind: one answer for the whole field

Field	Value	Records	Share
fcc_listed	"No"	41,753	98.9%
	"Yes"	472	1.1%

A yes/no field, 98.9% one answer. It is not broken, but it carries almost no information — unless what you want is precisely those 472.

The corrected table

Field	Apparent coverage	After stripping boilerplate
notes	100.0%	20.9%
toxicity_oral	100.0%	12.3%
toxicity_dermal	100.0%	6.2%
toxicity_inhalation	100.0%	1.7%
assay	64.5%	10.5%
fda_unii	71.9%	25.1%
hmdb	55.9%	17.4%
chembl	13.8%	0.0%
chebi	11.0%	0.0%
fcc_listed	100.0%	1.1%
occurrence	100.0%	51.4%
formula	51.9%	51.2%

The last row is the control: formula has 295 records reading "unspecified" and barely moves. Not every field has this problem.

Three ways to fill a field with nothing

1. A placeholder — Not determined, None found. It says outright that there is nothing.
2. A default template — "95.00 to 100.00". It looks like a number, and a plausible one. This is the dangerous kind.
3. Mis-scraped interface text — View, Search. Not even a number.

The second is the hardest to catch because it passes every shape check: a valid range, correctly ordered, physically plausible. None of the three checks that need no chemistry would find it.

There is one way to catch it: count how many distinct values the field holds.

In practice

1. Before quoting a field's coverage, count its distinct values. If the most common one accounts for four fifths, that coverage figure is false.
2. Read the assay field alongside the supplier count.
3. Ignore chembl and chebi.
4. This sweep is cheap — one GROUP BY — and it caught something a hundred articles had not.

What this doesn't establish

• I only swept strings under 45 characters occurring 200 or more times. Longer boilerplate is missed.
• "95.00 to 100.00 is a default" is an inference. It is possible that twenty-odd thousand materials genuinely specify that range — but 97.5% of zero-supplier records having it makes me disbelieve that.
• I did not check the source site, so "View is mis-scraped button text" is inferred from the pattern.
• The corrected coverage figures mean "not boilerplate", not "correct". A cell containing a real number does not make that number right.
• This is a September 2026 snapshot.