Retro-aligned ground truth for the Leibniz Nachlass: alignment prototype, factory yield, and a preliminary audit (Leibniz Legible, Phases B2 to C2)

Abstract

Three reports in sequence on minting training data for Leibniz's hand by retro-alignment: projecting the reading text of copyright-expired Akademie-Ausgabe volumes (§70 UrhG, 25-year term for scientific editions) onto machine-transcribed manuscript lines and keeping the pairs that match, the method the Bullinger correspondence project used at scale. Part 1, the prototype gate: measured on the PHILIUMM validation split under real HTR noise, the aligner mints 98.8% of lines at 97.5% precision on favourable material, with zero false mints of edition-omitted lines in every omission condition. Part 2, the factory at scale: 21 of 31 expired volumes read from free digital copies of the print (6,188 printed pieces, 26.2 million characters of reading text), 17,162 catalogue piece citations of which 11,595 localize to exact canvases through the library's IIIF folio labels, and a mint of 297,424 open-licence ground-truth lines, 5.9 times the target, stratified by fair copy, light revision, heavy revision and scrap with a per-stratum threshold.

Part 3, the audit of that mint, is unfinished and is published as such: 20 of the 200 sheet lines were judged by a non-specialist; five of the six non-correct verdicts sit on lines where the machine reading agrees with the minted text at 0.82 to 0.98 similarity, so the verdicts are recorded as preliminary and the precision gate is deferred to the fine-tuning ablation of the next phase. None of Part 3 is a pass. Only the constituted reading text of expired volumes is used; editors' introductions, apparatus and commentary never enter the pipeline.