Dataset · revision 2026-10-23.5
Cross-register match precision benchmark
1,001 reviewed links between regulatory records and PharmaTek organisation and site profiles (DMF holders, CEP holders, GMP documents, FDA enforcement events and FDA sites), before and after the 23 October 2026 link repair, plus one link for every distinct non-exact DMF holder and enforcement name. Each link is judged as the same legal entity or site, a related entity in the same group, wrong, or unclear.
Revision 2026-10-23.5, published 23 Oct 2026, 15:36 UTC · SHA-256 5e3504de9af9d22ef30cad66b9e390d829a858c157fb03d673f0972d89494521
- Population
- All stored links for six relations. Each sample takes 25 links from the exact-name stratum and 35 from the non-exact stratum (reproducible ordering), so the riskier stratum is over-sampled; estimates are weighted back by stratum size. Post-repair samples exclude every link reviewed earlier.
- Method
- Each link was reviewed record by record by PharmaTek against the source record detail (certificate and substance, document address, event or site address) and the linked profile. Estimates weight each stratum's share correct by the stratum size (stratum_size); the samples are not pooled, because the strata have very different sampling fractions. The 95% lower bound combines one-sided exact (Clopper-Pearson) 97.5% bounds per stratum with the same weights, which holds at 95% or better overall; the calculation is below. Rows from rule-tuning samples informed the final rules and are excluded from the estimates. All current verdicts are internal; a blind second review of 300 further links is pending.
- Limitations
- Samples are small: with 60 links reviewed, zero errors is still consistent with an error rate of up to about 5% (95% upper bound). Review was internal, not independent. Stricter rules trade recall for precision: some correct links are no longer made.
- Licence
- PharmaTek's review verdicts are released under CC BY 4.0 (credit “PharmaTek Intelligence”). Record names come from the registers listed, with their own terms; CEP holder names are reproduced from EDQM with permission.
Columns
| Column | Meaning |
|---|---|
| benchmark_row_id | Row ID in PharmaTek's benchmark table (from revision 2026-10-23.5) |
| edition | 2026-10-23: sample before the repair; 2026-10-23-post-repair: fresh sample after it; 2026-10-23-validation: test of the page-level site guard; 2026-10-23-distinct-names: one link per distinct non-exact name; 2026-10-23-cep-v2: fresh CEP sample after the second CEP fix; *-rule-tuning: rows used to tighten a rule (not used for the estimates) |
| relation | Which link was reviewed (for example DMF holder to company, GMP document to site) |
| stratum | exact or non_exact name match, used to weight the estimate; validation rows: guard_pass or guard_drop; non_exact_distinct: one row per distinct name, not weighted |
| stratum_size | Number of links of that relation and stratum when the sample was drawn (the stratum weight; repeated on each row, do not sum) |
| source_record_id | PharmaTek ID of the regulator record (DMF, CEP, GMP document, enforcement event or FDA site) |
| source_record_name | Organisation or site name as it appears in the regulator's record |
| source_record_detail | What the verdict was checked against: certificate number and substance, document address, event, or site address |
| linked_entity_id | PharmaTek ID of the company or site profile the record was linked to |
| linked_entity_name | The PharmaTek company or site profile the record was linked to |
| linked_entity_detail | Country or address of the linked profile |
| verdict | correct (same legal entity or site; renamed entities count as correct), group (a parent, subsidiary or sister entity), wrong, or unclear |
| review_note | Reason recorded for every verdict other than a plain match (group, wrong, unclear, renamed); blank where name, form and country agree |
| reviewer | Who reviewed: all current verdicts are PharmaTek's internal review, not independent |
| reviewed_at | Date of review |
Query used for this release
-- Estimate and stratified 95% lower bound for one relation, from the published CSV loaded as table b.
-- Sampling: within each relation, a seeded random sample of 25 links from the exact-name stratum and 35
-- from the non-exact stratum, drawn from the stored links on the date shown; stratum_size is the frame.
with s as (
select relation, stratum, max(stratum_size) as n_h, count(*) as k,
count(*) filter (where verdict = 'correct') as ok
from b where edition = '2026-10-23-post-repair' -- CEP: '2026-10-23-cep-v2'
group by 1, 2
), w as (
select *, n_h::numeric / sum(n_h) over (partition by relation) as weight from s
)
select relation,
sum(weight * ok / k) as weighted_estimate,
-- one-sided exact (Clopper-Pearson) bound per stratum at 97.5%; with ok = k it is 0.025^(1/k)
sum(weight * case when ok = k then power(0.025, 1.0 / k) end) as lower_bound_95_if_all_correct
from w group by relation;
-- Strata with errors need the beta quantile: lower = BETA.INV(0.025, ok, k - ok + 1).Revisions
Every published revision stays downloadable. Cite the revision you used; a later revision supersedes it but does not replace it.
| Revision | Published | Rows | SHA-256 |
|---|---|---|---|
| 2026-10-23.5current | 23 Oct 2026, 15:36 UTC | 1,001 | 5e3504de9af9d22e… |
| 2026-10-23.4 | 23 Oct 2026, 12:06 UTC | 1,001 | 746bfbb1cff3a460… |
| 2026-10-23.3 | 23 Oct 2026, 11:40 UTC | 941 | c4e170c0ee5a2179… |
| 2026-10-23.2 | 23 Oct 2026, 10:55 UTC | 890 | 656eaa50556d7166… |
| 2026-10-23.1 | 23 Oct 2026, 02:17 UTC | 410 | ac52e69985257065… |
Earlier revisions were published at a dated path that was overwritten when a correction was made; they are restored here from version control, timed by the change that published each one.
Sources
- FDA Drug Master File list
- EDQM Certification Online Database
- EudraGMDP (European Medicines Agency)
- MHRA GMDP (reused with attribution; OGL v3 assumed, confirmation requested)
- FDA Drug Establishments Current Registration Site
- openFDA drug enforcement and FDA Warning Letters
Certification Online Database. European Directorate for the Quality of Medicines & HealthCare (EDQM), Council of Europe, (2026), URL: https://extranet.edqm.eu/publications/recherches_CEP.shtml. Accessed 19 October 2026. Reproduced with permission; the EDQM and the Council of Europe accept no liability for any consequence arising from its use.
Cite as: PharmaTek Intelligence, “Cross-register match precision benchmark”, revision 2026-10-23.5 (SHA-256 5e3504de9af9), https://pharmatek.co.uk/data/match-precision-benchmark. See the methodology and corrections.