1[2 {"id": "S1", "source": "arxiv-2410.12944v3", "short": "Paradis et al. (2024), Google enterprise RCT, 96 developers, one complex coding task",3 "extract": "the coefficient on the experimental condition (ExpCon) in Model 2 of Table III (logged time on task, developer-level covariates), its standard error, and the 95% confidence interval and p-value quoted for it in Section VI-C"},4 {"id": "S2", "source": "arxiv-2507.09089v2", "short": "Becker et al. (2025), METR field RCT, experienced open-source developers, real issues",5 "extract": "the headline regression-based estimate of the percentage change in issue completion time when AI is allowed, as stated in Section 3 (with the estimator defined in Sections 2.3 and D.1-D.3), together with whatever uncertainty is printed for it in the extract"},6 {"id": "S3", "source": "arxiv-2411.01067v2", "short": "Bono and Xu (2024), Microsoft Security Copilot RCT, IT administrators",7 "extract": "the aggregate 'All Content' row of Table 3 (task duration in minutes, holding accuracy constant): the treatment and control durations and the printed improvement, together with whatever uncertainty is printed for it"},8 {"id": "S4", "source": "arxiv-2511.13860v1", "short": "Bono (2025), Microsoft phishing-triage agent RCT, security analysts",9 "extract": "the 'Full Corpus' column of Table 3 (time per email): the Aware and Control times, the printed % time savings and its p-value"},10 {"id": "S5", "source": "arxiv-2304.09339v1", "short": "Kreitmeir and Raschky (2023), difference-in-differences around Italy's ChatGPT ban, GitHub users",11 "extract": "the Treat x Post coefficient and standard error for the Output measure (column 3) in panel A of Table 1 (users with affiliation), and its sample size"}12]13