How we assess a perfume combination
Layer or Not? compares published perfume profiles and displays visitors’ reported experiences alongside that comparison. This English guide explains the current method; the full technical report is available in Spanish.
1. What the two scores mean
The model’s 0–10 score describes relative documentary fit against other catalog pairs. The separate 1–5-star community average describes reported enjoyment from visitors who say they tried the combination.
An 8 is not an 80% chance of success, and a low score does not prove a combination smells unpleasant. Widening the score range does not demonstrate better prediction of real-world enjoyment.
2. Identity, notes and source coverage
The imported collection contains 7,278 profiles. Brand, name and known concentration identify a version: an unknown concentration does not automatically inherit another edition’s data. Aliases are normalized and repeated note labels are removed.
There are 830 profiles with note lists from several publishers and 301 with a manufacturer reference. The current source review includes 12 manufacturer checks and applies 9 complete official pyramids. Source lists stay separate; they are not pooled to invent a larger pyramid. Coverage is partial, and different publishers may reuse the same original information.
3. Shared and complementary notes
Identical normalized notes have relationship strength 1. Different notes receive support when they co-occur in at least 5 individual perfume profiles in a corpus of 6,495 profiles. The corpus keeps one profile per brand and name, prioritizing manufacturer information when available.
nAB counts profiles containing both notes; nA and nB count each note’s occurrence. This regularized Dice association preserves its magnitude. It is not converted into a per-association percentile. Co-occurrence in one perfume is supporting catalog evidence, not an experiment with two perfumes layered on skin.
4. Bilateral coverage and note weights
bilateral relationship = 2 × Σ(association × smaller matched-note weight) / Σ(weights of both profiles)
Bounded inverse document frequency reduces the dominance of ubiquitous labels; a note with no known frequency has weight 1. A maximum-weight one-to-one assignment selects links without reusing a note. The denominator includes both profiles, including unmatched notes, so a short list cannot simply hide the unmatched part of a longer list. Note rarity still does not measure perceived intensity or formula concentration.
5. Families, stages and component weights
The initial weights are 50% note relationships, 35% family distributions and 15% published stages. A family comparison adds the smaller share of each classified family and requires at least 60% classified-note coverage in each profile. Stages compare declared top, heart and base positions, with at least two comparable stages required.
Unavailable components are excluded and the remaining weights are renormalized. Families and stages reuse note information; they are not independent sensory confirmations. The interface also shows alternative editorial weightings of 65/25/10 and 35/45/20 against the same fixed reference. Their range is a sensitivity check, not an experimental confidence interval. These weights are documented design choices, not coefficients validated by the cited research.
6. The relative 0–10 scale
The weighted raw index is positioned against 10,000 distinct, uniformly sampled catalog pairs drawn from 5,439 eligible versions with known identity and at least three notes. The reference is frozen by version, using seed 20261025. Ties use their midpoint rank; the percentile is multiplied by ten and rounded to half-points.
Zero support remains 0. A score of 10 is reserved for a raw index of at least 9.5; other cases are capped at 9.5. The wider range follows from this comparative scale. Updating the catalog requires rebuilding and versioning its reference, so relative scores can move even when a particular pair’s notes remain unchanged.
7. Provisional, unavailable and on-skin results
A usable note profile on both sides allows a comparison. An unknown concentration or fewer than three distinct notes on either side makes the score provisional. If either perfume lacks usable notes, the score is unavailable: the app offers alternative profiles and source links rather than inventing missing data. Missing stages are not filled in.
Commercial note lists do not reveal full formulas, doses, measured projection or evaporation curves. Concentration labels do not receive an automatic score bonus. A 1:1 starting suggestion is a way to compare an initial trial, not an optimized ratio. Record quantities, surface, elapsed time and any dominance, and compare the combination with wearing each perfume separately.
8. Community ratings stay separate
The visible star average uses submitted ratings and shows the real vote count. It does not change the model score. Visitors can update one rating per browser and pair; a browser is not proof of an independent person, and the declared trial is not independently verified.
This smoothing is equivalent to one prior count in each of the five star categories, a uniform Dirichlet prior. Those are not five real participants. It affects community ordering, not the displayed mean or real count. Optional surface, time, balance, preference and spray-count summaries use only responses to that particular field. Unanswered questions and older reviews do not acquire invented conditions. Comments are not turned into synthetic numeric ratings.
9. Reproducible audit and stability
The audit evaluates another 10,000 pairs using seed 20261026, excluding all reference pairs. The previous and current methods are compared on the same pairs. This tests score distribution and computational behavior, not whether the mixtures are pleasant.
| Measure | Previous method | Current method |
|---|---|---|
| Pairs evaluated | 10,000 | 10,000 |
| Median | 8 | 5 |
| Mean | 7.78 | 5.01 |
| Standard deviation | 1.76 | 2.87 |
| 10th percentile | 5.5 | 1 |
| 90th percentile | 9.5 | 9 |
| Scores of 8 or higher (%) | 60.5 | 22.71 |
The share scoring at least 8 changed from 60.5% to 22.71%. That is a property of the scoring scale and available descriptors, not the proportion of enjoyable perfume combinations.
| Pair | Previous | Current |
|---|---|---|
| Dior Sauvage Eau de Parfum + Maison Francis Kurkdjian Baccarat Rouge 540 | 9.5 | 5 |
| Tom Ford Oud Wood + Maison Francis Kurkdjian Baccarat Rouge 540 | 9.5 | 7.5 |
| Maison Francis Kurkdjian Baccarat Rouge 540 + Parfums de Marly Layton | 9.5 | 3 |
| Maison Francis Kurkdjian Baccarat Rouge 540 + French Avenue Liquid Brun | 9.5 | 6.5 |
A stability check resampled the 10,000-value reference 20 times and evaluated 10,000 audit pairs, using seed 20261027. 93.97% of comparisons kept exactly the same score; the maximum change was 0.5 points. This measures reference-sampling sensitivity, not uncertainty about human preference.
Download the audit JSONFull technical audit (Español)
Reproduce the repository outputs with scripts/build-harmony-corpus.mjs, scripts/build-note-frequency.mjs, scripts/build-score-reference.mjs, scripts/audit-scores.mjs and scripts/audit-score-stability.mjs. This English page is regenerated by node scripts/build-method-en.mjs from the same local data and audit files.
10. Research, limitations and next validation
The research motivates recording intensity, dose, body surface, elapsed time, dominance and personal variation. It distinguishes similarity and discrimination from enjoyment. None of these studies validates the editorial weights or predicts the outcome of a particular pair of commercial perfumes from its marketing notes.
- Lapid, Harel and Sobel (2008) — Relative intensities and pleasantness in binary mixtures.
- Ma et al. (2021) — Intensity, pleasantness and component perception in controlled mixtures.
- Ma et al. (2023) — Perceptual interactions in complex aroma mixtures.
- Lenochová et al. (2012) — Individual responses to perfume and body odor.
- Ravia et al. (2020) — Perceptual similarity of odor mixtures.
- Dhurandhar et al. (2023) — Descriptors and odor discrimination.
- Chang et al., eLife — Learning in mixture discrimination.
- Keller et al. (2017) — Olfactory prediction evaluated on held-out human measurements.
- Stan, official statistical documentation — Dirichlet distributions.
- Scento, CC BY 4.0 — Materials and aliases used to normalize note names.
To evaluate predictive accuracy, future trials need measured outcomes for specific pairs, comparison with wearing each perfume separately, repeated observations and an evaluation set excluded from parameter tuning. Sampling should include low and middle score bands as well as high ones. Voluntary community feedback can be biased; the current documentary index has not been validated as a predictor of layering enjoyment on skin.