Beginner perfumer #120: the evaluation concentration really answers one yes-or-no question

About a 5 minute read.

The short version

• Two fields carry a concentration: odor_strength (recommends N% or less) and odor_descriptions (the concentration actually used in evaluation)
• Across the 797 materials with both, 94.1% carry the same number
• Their coverage differs sharply: odor_strength on 802 records, odor_descriptions on 7,620
• Recomputed on the 604 materials resolvable to the corpus:

Evaluation concentration	Materials	Median dose
100%	377	2.50%
10%	151	1.00%
1%	65	0.86%
0.1%	6	0.85%

• rho = +0.438, and dropping the 100% group takes it to +0.114
• This also corrects the 0.679 I published in #107

Two fields carrying one number

Last round's 7-methyl coumarin turned up a contradiction: its strength rank reads 1 (weak) while the recommendation in the same field reads 1% or less.

Following that up showed this database writes concentration in two places:

Field	Content	Records
odor_strength	medium ,recommend smelling in a 10.00 % solution or less	802
odor_descriptions	at 100.00 %. sweet woody rose...	7,620

The first is a recommendation (how far to dilute before smelling). The second is a record (what the evaluator actually used).

Across the 797 with both, 750 (94.1%) carry the same number.

But one of them only exists when dilution is needed

The difference is not in the values; it is in when the field gets filled.

Cross-tabulating them (rows: what was actually used; columns: what is recommended):

Actual ＼ Recommended	0.01%	0.1%	1%	10%	50%
0.1%	3	38	2	1	0
1%	0	5	188	1	0
10%	0	0	5	504	0
50%	0	0	0	0	11
100%	1	1	5	21	1

The diagonal is clean. But look at the bottom row: only 29 records were evaluated at 100%.

Meanwhile odor_descriptions carries an evaluation concentration on 7,620 records across the database, a large share of them at 100%.

The odor_strength field is filled almost exclusively when the answer is "dilute it". Whether it has a value carries more information than what that value says.

The dose ladder, recomputed

Using the name resolver built over three rounds to match the corpus back to the database gives 604 materials appearing in 3 or more formulas and carrying an evaluation concentration:

Evaluation concentration	Materials	Median dose
100%	377	2.50%
10%	151	1.00%
1%	65	0.86%
0.1%	6	0.85%

The 100% cell reads 2.50%, and the other three all sit between 0.85% and 1.00%.

The correlations:

Sample	n	rho
All	604	+0.438
Excluding the 100% group	227	+0.114
Only 10% and below	225	+0.121

Remove one group and the correlation collapses.

So the field answers a yes-or-no question

Four values (100%, 10%, 1%, 0.1%) look like a four-step ladder. What they actually answer is one question:

Can this material be smelled neat, or does it have to be diluted first?

Materials you can smell neat have a median dose of 2.50%. Materials that need diluting have a median dose between 0.85% and 1.00%, whether they are diluted to 10%, 1% or 0.1%.

A hundredfold difference in dilution buys 0.15 percentage points of dose.

I caught half of this in #107, where I wrote that a hundredfold drop in evaluation concentration produces only a 5.2-fold drop in dose. This round's numbers are more extreme: not 5.2-fold, but almost nothing.

And a correction to one of my own numbers

In #107 I published a correlation of rho = 0.679. This round it comes out at 0.438.

The difference is the sample:

	#107	This round
Corpus threshold	20 or more formulas	3 or more
Name matching	name + head_synonym	the full resolver built over three rounds
Materials matched	117	604

Both numbers are correct for their samples, and the 604 one is more trustworthy. The 117-material set only included materials appearing in 20 or more formulas, which is itself a selection biased toward frequently used materials.

The direction is unchanged (positive, higher evaluation concentration means higher dose) and the strength is about a third lower. What matters more is that the correlation comes almost entirely from one binary distinction rather than from a ladder.

What this is good for

1. at 100.00 % means you can smell it neat, and it usually goes in heavier (median 2.50%)
2. Any number below 100% means dilute before smelling, and start your dose around 1%
3. Do not infer dose from the difference between 1% and 10%. Those two cells differ by 0.14 points
4. Whether odor_strength has a value is itself the signal. If it does, the material needs diluting

(In 2025 Wachowiak and colleagues argued in J Neurosci that olfactory research's concentration ranges do not line up with natural ones (PMID 40044450), and Cometto-Muñiz and Abraham measured concentration-response functions for homologous alcohols (PMID 18950650). Those are about noses. This is about how fields get filled in.)

What this doesn't establish

• The 604 came from a threshold of 3 or more corpus formulas. Raise it and the sample shrinks toward frequently used materials, and the correlation may rise again. That is exactly what happened in #107.
• The 0.1% cell holds 6 materials. Do not treat 0.85% as a stable estimate.
• I did not address the possibility that evaluation concentrations were themselves typed casually. I swept specification ranges for round endpoints but ran no equivalent check on evaluation concentrations.
• The 94.1% agreement comes from 797 records, the intersection where both fields have values, not from the database as a whole.
• "Median dose" is the corpus's, not the market's.
• Correcting #107's number does not make #107 wrong. Its core claim, that evaluation concentration predicts dose better than the strength rank, still holds on this sample. The effect is smaller than I reported and its source is simpler than I thought.