How a scorecard is computed
Every number on a supplier scorecard, with its definition, its denominator and its failure mode. If a figure on this site cannot be reproduced from this page plus the published dataset, that is a bug and the corrections log is where it gets fixed.
The general rule
Every statistic is computed over the records the index holds and nothing else. There is no adjustment for market share, no imputation of missing values, no exclusion of inconvenient records, and no smoothing. Where a figure cannot honestly be computed it is not computed: below n=3 the index prints the raw values instead of a summary, and between n=3 and n=8 the badge beside the figure turns amber.
All quantiles are type 7 — the R default — computed by linear interpolation between order statistics. Medians rather than means throughout, because the purity distribution is compressed at the top and has a long left tail, and a mean over that shape is dragged by exactly the observations you most want to see individually.
Median purity
median(purity) over every record for the supplier that carries a purity figure. Records for substances with no defined purity are excluded from the numerator and the denominator both, because a purity figure for them would be an invention rather than a measurement.
Failure mode. This figure is dominated by product mix. A supplier whose index coverage is mostly semaglutide will out-median one whose coverage is mostly short bioregulators, and neither fact concerns the supplier. The per-compound delta column on each scorecard exists because of this and is the comparison to use.
Interquartile range and the 10th percentile
Q3 − Q1 and quantile(purity, 0.10). The IQR answers "how consistent", the 10th percentile answers "how bad is a bad lot". Both are more informative than the median for a buyer, because a buyer receives one lot rather than the median of all lots. The 10th percentile is suppressed below n=10, where it would be an extreme order statistic dressed up as a percentile.
Pass rate against the index specification
count(purity ≥ compound.spec.purity) / count(records with a definable purity). The specification is per compound class and every threshold is stated with its source on the specifications pages. A pass rate is a comparison against a threshold the index chose; it is not a regulatory status and it is not a certification.
Reproducibility
median(max(purity) − min(purity)) over the supplier's batches with three or more independent assays. This is the statistic the batch-first data model exists to make honest: repeated assays of one physical lot are repeated draws around one latent value, so the spread is a real property of the material and the fill rather than a summary of unrelated lots.
Failure mode. It requires repeats, and repeats are rare — most lots are tested once. Suppliers with few multiply-tested lots get no figure at all, and the absence of a reproducibility number is not a bad reproducibility number.
COA divergence
coa.purity − median(independent purity for the same lot), computed per batch and then aggregated as a median over batches. Positive means the certificate reads higher than the measurement.
The index reports both the median divergence and the share of divergences that run in the supplier's favour, and the second matters more than the first. Two measurements of the same thing disagree; that is scatter. Seven out of seven disagreeing in the same direction is not scatter, and the scorecard says so in those terms.
Documentation completeness
A 100-point rubric applied to the certificates the index holds for that supplier. It scores what the document contains, never how it looks.
| Component | Weight | Why it is worth that |
|---|---|---|
| A purity figure with a stated method — column, gradient, wavelength, reporting threshold | 25 | A purity figure without a method cannot be compared with any other purity figure, including a re-test of the same lot. |
| Named related substances with retention times, rather than a single total | 25 | Named impurities are what make an integration checkable and what let two lots be identified as the same bulk. |
| Peptide content, stated with its determination method | 20 | Content is what decides how much peptide is in the vial. Purity does not. |
| Water and counter-ion, each with a method | 15 | Together they close the gravimetric mass balance. Without them the balance cannot be closed at all. |
| Endotoxin, with the assay type | 15 | The one property no chromatogram bears on. |
Provenance breakdown
Records are classified buyer-sourced, supplier-supplied or lab-retained. Across the 88 lots in the index where both a supplier-supplied and an independently purchased sample of the same lot were assayed, supplier-supplied results read a mean 0.31 pp higher (95 % CI 0.19 to 0.43). Each scorecard therefore reports the buyer-sourced subset separately, and rule AX-R10 flags the individual lots where the gap exceeds 0.8 pp.
What is deliberately absent
- No composite score. A single number requires weights; weights are an editorial opinion; and a reader would treat the result as a measurement. Thirteen columns and a sort control are more honest and no less usable.
- No price, no delivery time, no customer-service assessment. The index measures vials. It has no data on any of those and will not gesture at them.
- No supplier ranking. The suppliers are ordered by index coverage on every table, which is a property of this dataset rather than a claim about them.
- No adjustment for supplier co-operation. Suppliers who send samples are not rewarded for it, and their samples are down-weighted rather than up-weighted.
Ratings and reviews
The rating shown on a scorecard is the aggregate held for that supplier across this index's contributor network, and the three reviews beneath it are signed statements from named contributors about material they tested themselves. Neither feeds any statistic on this site. They are published beside the arithmetic, not inside it, and where a review contradicts the arithmetic both are left standing.