Same thesis, two languages: LLM-associated vocabulary in the English and Portuguese abstracts of 914,000 Brazilian theses, 2013-2024

Studies of post-ChatGPT ``excess vocabulary'' mostly measure English papers; Korean journal abstracts and dissertations at one Brazilian university have also been measured (Section 1). In Brazil's national thesis catalogue most records carry a Portuguese abstract (resumo) and an English one, sharing topic, author, programme and year but not language: a paired design. Using the CAPES catalogue (913,707 theses and dissertations with both abstracts valid, 2013-2024, 729 institutions), this paper measures pre-registered LLM-associated style words in English abstracts and their Portuguese translations in resumos. Between 2020-2022 and 2024 the English marker rate rose by a factor of 2.56 (95% cluster-bootstrap interval 2.49-2.64) and the Portuguese rate by 1.75 (1.70-1.79). The registered primary estimate, the difference in log changes, is 0.383 log points (0.367-0.401), a 47% larger relative rise in English. Neither registered placebo reaches its flag threshold: control words drift by 0.016 log points (0.013-0.018), 4% of the primary estimate, and a fake break in 2021 gives -0.010 (-0.025 to 0.005). The point estimate is positive in all nine broad areas and all four degree types; the professional doctorate (89 theses before; 0.553, 0.168-0.788) was first dropped by an unregistered size rule. The Portuguese resumo is not an untouched baseline: its marker rate rose by 75%, and an exploratory excess-vocabulary analysis finds Portuguese words such as crucial (7.3 times the expected 2024 frequency), destacando (3.5) and promovendo (3.2), as Amarante et al. (2026) report at one university; the national data replicate and extend that finding rather than establish it. The gap depends on the marker list: with the 291-word English list of Kobak et al. it shrinks to 0.088 (0.072-0.106). The design is observational and detects a lexical shift, not LLM use as such.