Beginner perfumer · 79
The official blending list is truncated — capped at 120, and 55.5% of materials sit exactly at the cap
· 10 min read
The database's blending suggestions run to 673,662 rows, which looks like a complete relationship map. It is not. Each material's suggestions split into odour and flavour, each capped at 60, for a total ceiling of 120 — and of the 7,129 materials with suggestions, 3,957 (55.5%) stop exactly at 120, with none above it. So for more than half of all materials you are seeing a truncated top sixty, and nothing in the data tells you what was cut. That also explains two other things: of the 344,065 directed pairs that map back to material IDs, only 18.9% are mutual; and the 69% of empirically co-occurring pairs I previously found absent from the official list.
About 4 minutes.
The "blends well with" list on a material page looks like a complete relationship map. There are 673,662 rows of it database-wide, a number large enough to suggest it covers everything.
It is a ranking, and it is truncated.
The cap is 120
7,129 materials carry blending suggestions. Counting them per material:
| Suggestions | Materials | Share |
|---|---|---|
| Exactly 120 | 3,957 | 55.5% |
| 111 | 51 | 0.7% |
| 106 | 48 | 0.7% |
| 98 | 47 | 0.7% |
| Everything else | … | … |
Materials above 120: zero.
Break it down further and the suggestions split into two kind values — odour (336,574 rows) and flavour (337,088 rows). The maximum for each is 60.
60 + 60 = 120. A hard ceiling.
So for more than half of all materials, what you see is a top sixty, not the whole set. And nothing in the data records what was cut.
This explains the 81% asymmetry
Mapping the suggestion names back to material IDs gives 344,065 directed pairs (another 203,516 rows have names with no exact match in the materials table, mostly compounded products and trade names).
Then a simple question: if A suggests B, does B suggest A?
| Pairs | Share | |
|---|---|---|
| Mutual (A→B and B→A) | 65,035 | 18.9% |
| One-way | 279,030 | 81.1% |
Four fifths run one way.
And the cap accounts for most of that. A popular material appears in many other materials' top sixty, but its own list has only sixty slots — not enough to hold everyone who listed it.
This is not a data error; it is what rankings do. Any "top N per item" list generates this asymmetry.
It also explains the 69% I calculated earlier
In I bought thirty materials and got stuck I computed which materials co-occur across 954 published formulas and compared them against the official suggestion lists, finding that 69% of the empirically co-occurring pairs were not on the official list — including obvious combinations like oakmoss with patchouli and bergamot with lavender. The two lists a single material carries also point in different directions.
At the time I treated that as two data sources seeing different things.
There is now a simpler explanation: the list has sixty slots.
Oakmoss with patchouli being absent does not mean the database regards the combination as bad. It may just have ranked sixty-first.
My conclusion then was pointed the right way — do not treat the official list as a complete table of workable combinations — but I had not found the reason. Now I have.
What to do
- Absence from the list does not mean they clash. For 55.5% of materials the list could not have fitted them.
- Presence on the list is a meaningful signal. A combination that makes the top sixty has at least been explicitly recommended.
- Two materials each listing the other is a stronger signal. Only 18.9% of pairs manage it.
- Do not judge a material's versatility by list length. For most materials, the length is just the cap.
- To find combinations that were not listed, use a formula corpus rather than the suggestion list. Their biases differ.
What this doesn't establish
- I do not know how the sixty are ranked. There is no score or rank column, so "top sixty" is inferred from the cap, and I cannot verify the ordering criterion.
- The 203,516 rows that do not map back to IDs are excluded from the symmetry calculation. Including them would change the proportion.
- Symmetry was computed on exact name matches. A material written differently in two places counts as unmatched.
- "The cap explains most of the asymmetry" is an inference. Testing it needs the ranking criterion and the untruncated lists, and I have neither.
- The 69% figure comes from a 954-formula corpus with its own bias — publicly shared formulas are not all formulas.
- I did not evaluate the quality of any blending suggestion. This article is about the shape of the list, not its contents.