Partial computational replication and preprocessing sensitivity in AI assisted email composition

Computational replication can distinguish reproducible statistical results from uncertainty about the data transformations that produced them. This article reports a partial computational reimplementation of Buschek, Zürn, and Eiband’s study of parallel phrase suggestions in email composition. Public, versioned data and read-only reference code were used to reproduce selected task-duration, suggestion-selection-time, and accepted-and-retained suggestion analyses. The selected task-duration core coefficients and four acceptance coefficients matched the precision reported in the original paper. Some selection-time quantities and task-duration post-hoc comparisons remained unmatched. A source audit identified a public denominator transformation that removed 3,789 displayed-list trials from 71 zero-retention tasks. A post hoc sensitivity analysis, specified after the source audit and before fitting, retained the model specification and success definition used in the faithful reimplementation while restoring directly recorded display counts. It expanded the acceptance analysis from 149 participants and 397 tasks to all 156 participants in the cleaned study sample and 468 suggestion-enabled tasks. The two suggestion-count contrasts among native speakers (three or six versus one) and their language-group interactions remained positive, with all four profile intervals above zero. An initial standard-error agreement check failed; a transparently documented post hoc Richardson-Hessian diagnostic subsequently supported numerical stability, while the original failed flag was preserved. These results document the reproducibility of selected findings and their sensitivity to a specific preprocessing decision. They do not reconstruct the complete original pipeline or validate broader claims about language, culture, or workplace productivity. AI assistance disclosureAn OpenAI assistant substantially assisted literature and source retrieval, extraction, research planning, data processing, Python and R code generation and revision, tool-mediated execution and debugging, numerical diagnosis, interpretation checks, and manuscript drafting and editing. AI-assisted reviews of mapping and mathematical implementation were also used. The human author is responsible for research judgments, review and verification, and approval of the final manuscript, and has reported completing review of the manuscript and relevant materials. This author self-report does not independently certify a line-by-line code audit or an author-executed rerun. The precise historical model identifier is not reliably recorded and is not guessed. The results arise from documented computations on existing public research data; no synthetic observations are presented as participant data. The full scope, unsuccessful checks, diagnostic amendments, execution deviation, later rerun differences and limitations are retained in the manuscript and supporting record. The assistant is not an author or record contributor. Data sources and rightsOriginal study: Buschek, Zürn and Eiband (2021), doi:10.1145/3411764.3445372. Original data: https://osf.io/7q4c8/. Participant data, identifiers and email text are not redistributed. The code retrieves four fixed original OSF revisions and enforces the SHA-256 values recorded in the article Appendix A and code/source_manifest.json. New-code MIT and rights-eligible original article/documentation CC BY 4.0 grants do not relicense original data, source excerpts, third-party material or underlying facts. LicensesArticle PDF and rights-eligible original documentation/aggregate expression: CC BY 4.0. New implementation contributions: MIT under code/LICENSE and code/LICENSE_SCOPE.json. Third-party notices remain in code/THIRD_PARTY_NOTICES.md. This is a component-specific mixed-license deposit, not a blanket CC BY or MIT grant over every ZIP member. Metadata are public and reusable under CC0 under Zenodo policy. Reproduction instructionsDownload the separate R2_reproducibility_instructions_v0.2.md file and R2_code_materials_v0.2.zip from this record. The archive includes frozen plans, historical failed checks, amended diagnostics, current aggregate verification and explicit historical/current comparisons. This is a preprint; no peer-review outcome or whole-paper reproduction is claimed. Direct reproducibility instructions: https://zenodo.org/records/23187196/files/R2_reproducibility_instructions_v0.2.md?download=1