Trained Stanza models for the nine reported iterations of the human-in-the-loop bootstrapping experiment described in Correction as Annotation: Bootstrapping a Dependency Parser for Documentary Medieval Latin. The corpus is DALME-Marseille, consisting of 160 inventories drawn up in the city of Marseille between 1258 and 1446. No existing Latin treebank model is usable on it, the best labelled attachment score among five baselines being 0.62. Each iteration sₙ here is a complete four-model Stanza pipeline—tokenizer, tagger, lemmatizer, dependency parser—trained from scratch on the pool of hand-corrected sentences accumulated through round n. They are not fine-tuned from one another, so the nine together are the learning curve the paper reports: 204 to 1,804 sentences of in-domain annotation, 33.2 hours of correction, and a rise from 0.80 to 0.98 UPoS and 0.48 to 0.92 LAS against the out-of-domain model that seeded the loop. Also included are the three sets of 300-dimensional word vectors trained on the corpus itself, which the tagger and parser require at load time. Each iteration is bound to the set it was trained against: s₁–s₅ to v1, s₆–s₇ to v2, s₈–s₉ to v3. The five baseline Latin models are third-party Stanza releases and are not redistributed here. The corpus, the gold standard, the domain lexicon, and the analysis code are released separately at github.com/gpizzorno/correction-as-annotation. See the deposit README for full details.