BookKorpus

Methodology

Duplication KPI — definition v1.0

Version 1.0 · as of [YYYY-MM-DD]
Applies to every KPI figure quoted by BookKorpus. Figures always cite the definition version they were computed under. Changes are versioned and logged in §7.
§1The claim being scored

For a given lot, the duplication KPI states the share of titles for which no edition-level match was found in the public digital corpora listed in §2, as of the manifest date. It is computed and reported per lot — never as a blanket figure across our stock.

§2Corpora checked

Full-text repositories checked for a digitised copy:

  • HathiTrust
  • Google Books
  • Internet Archive
  • Open Library

National library catalogues (for the languages present in the lot) are used for edition identification and to resolve ambiguous matches. The exact catalogue set applied to a lot is stated in that lot's manifest.

§3Matching level

Matching is at edition level: same work, same edition or impression, identified by imprint, year and physical description. A different edition of the same work counts as a work-level match, which is reported alongside as a separate figure — it is never blended into the headline KPI.

Every match decision carries an evidence link on the manifest: the corpus record we matched against, or, for a no-match, the documented queries that were run.

§4As-of dating

Public corpora grow. Every manifest states the date on which its checks were run, and the KPI holds as of that date. Re-scoring an existing manifest against a later date can be agreed as part of a transaction.

§5Match error

No matching process is perfect. We publish false-negative and false-positive rates from manual validation on held-out samples: [validated ranges to follow].

Until a validation figure has been published for a corpus, no KPI relying on that corpus is quoted. No number appears anywhere on this site that has not passed validation; bracketed placeholders mark pending values.

§6Audit protocol

The KPI is designed to be checked by the buyer, from the manifest alone:

  1. Select any sample of rows from the delivered manifest.
  2. Follow the evidence links, or re-run the documented queries against the corpora named in §2.
  3. Compare the outcomes with the dispositions recorded on the manifest.
  4. Discrepancies beyond the published match-error band (§5) are handled under the guarantee terms of the lot.

A buyer's spot-check should reproduce our results. If it doesn't, that is our problem, not the buyer's.

§7Versioning

This definition is versioned; material changes increment the version and are logged here. Quoted figures always name the version they were computed under.

  • v1.0 — initial publication, [YYYY-MM-DD].
Questions

Queries on the definition, or on applying it to a specific lot: contact@bookkorpus.com. We reply within one business day.