Reliability: measurements and limits

Dated measurements describe the available evidence. They do not certify today’s marketed formula or product compliance.

Age of composition evidence

• Over 5 years : 154,660 / 628,782 (24.6 %)
• 2 to 5 years : 104,297 / 628,782 (16.6 %)
• Under 2 years : 116,415 / 628,782 (18.5 %)
• Unknown or unusable date : 253,410 / 628,782 (40.3 %)

Calculated on 2026-10-31 · Method

Dated official sources

• Known stated date : 7 / 100 (7.0 %)
• Unknown stated date : 93 / 100 (93.0 %)

Calculated on 2026-10-31 · Method

Consistency across sources

• Different lists : 11,862 / 51,630 (23.0 %)
• Identical lists : 39,768 / 51,630 (77.0 %)

Calculated on 2026-10-31 · Method

Additional distinct lists in the internal register : 17207 / 110808

Calculated on 2026-10-31 · Method

Name recognition in published cosmetic lists

• Strong method : 943,673 / 1,022,076 (92.3 %)
• Unrecognised name : 45,688 / 1,022,076 (4.5 %)
• Weak method : 32,715 / 1,022,076 (3.2 %)

Calculated on 2026-10-31 · Method

Published corrections

10 / 10

• 2026-10-23 : 10 / 10

Calculated on 2026-10-31 · Method

Sample review

Sample drawn on 2026-10-31, review in progress: 0 / 200 records reviewed.

Sampling seed : 20261031

Sampled records / published records in the stratum

• Cosmetics : 8 / 79959
• Supplements : 29 / 270505
• Official registers : 163 / 1532142

Calculated on 2026-10-31 · Method

Protocol and review file

Method : Age of composition evidence

One unit is a published record with an ingredient list, including partial lists. Families and historical records remain distinct and are not treated as unique commercial products. Dates describe a photo, a page observation or a label. Calendar thresholds match the photo rules: under two years, two to five years inclusive, then over five years. Consultation and administrative dates do not refresh the list. Future, invalid and missing dates remain unknown. Registers without a list are excluded from this denominator.

Method : Dated official sources

The denominator counts primary-source entries in the project source registry: administrations, legal texts, public registers and the public endocrine-list network. Language editions are separate entries, not independent authorities. Contributor databases, Wikidata, visitor submissions and the internal manufacturer-site aggregate are excluded. An entry is dated when its stated source-date field is usable. Consultation is separate. A missing field does not prove that the source provides no date elsewhere.

Method : Consistency across sources

A barcode enters this denominator when the internal register contains multiple distinct provenances (source, URL). Normalised list hashes are compared, including historical and unpublished observations. A difference may reflect a reformulation, variant or error; it establishes neither cause nor chronology. Divergences sum the distinct lists beyond the first for every barcode in the register, including single-provenance records.

Method : Name recognition in published cosmetic lists

Each name occurrence in published cosmetic lists counts once. Strong methods follow inci.STRONG_METHODS; another method with a linked name is weak; names without a match are unknown. This scope is not extrapolated to supplements or registers, whose recognition methods differ. Name recognition does not prove exact transcription.

Method : Sample review

The sample allocates 200 records proportionally across three strata using largest remainders. Rank is SHA-256(seed, family, identifier), without replacement and independent of input order. The CSV and manifest fix the cohort and sources. A person must check identity, ingredients verbatim, supplement doses and displayed regulatory status, then record juste, erreur mineure or erreur critique, a comment and review date. Missing or unreadable evidence remains pending. Rates and 95% Wilson intervals cover reviewed rows only and appear from 50 reviews. Partial review may be biased and does not establish a database-wide error rate; Wilson corrects neither that bias nor stratification. No independent audit is claimed. The scope covers native imported product records, including historical and listless records; alert-only and distributor-only identities without composition are excluded. The CSV source_locator field distinguishes a source record from a collective dataset that must be filtered by native identifier.

The CSV has no reviewer identity field. Juste requires every applicable check to pass; erreur mineure means a discrepancy without changing identity, ingredients, dose or status; erreur critique means a discrepancy affecting one of those points. Without access to evidence, leave the result empty and record the blocker in the comment. Calculations are fixed at build time; the page and API read the same table.

Method