The reproducible package

Every finished run produces a reproducible package: a single zip file that documents the analysis completely and lets a statistician re-run it. You download it from the run page once the results are confirmed - see Export results and the reproducible package.

A README.txt at the top of the package lists what is inside that particular package.

Always included

• The analysis plan - the research question, the chosen test and why, the variables and the assumptions.
• The results - the statistics, the effect size and the assumption checks, both readable and machine-readable.
• The results table - as Markdown and as CSV, plus an Excel workbook with every table on its own sheet.
• The methods - a statistical methods section written from what the run actually did, with no AI involved, and the references it cites in a file for a reference manager.
• The audit log - every step of the run, including each de-identified prompt that was sent to the AI model. An AI step that was not requested is recorded as not requested.
• The run specification - the provenance of the run and the reusable analysis settings.
• Session information - the R version and the version of every package that produced the result.
• The lockfile - the exact pinned package set the engine ran on.
• The analysis script - an R script that re-runs this exact analysis from the package folder.

Included when it applies

• The data - a copy of the dataset the analysis ran on. See Handle it like the data.
• A post-hoc table, for tests with pairwise comparisons.
• Table 1 - baseline characteristics by group.
• Figures - aggregate figures appropriate to the method, such as box plots, Q-Q plots, Kaplan-Meier curves, forest plots or ROC curves, with a list of captions.
• Warning flags - the rule-based cautions for this run.
• A plain-language summary computed by fixed rules, with no AI.
• The AI interpretation, when interpretation was on for the run.
• A manuscript draft (Markdown, Word and LaTeX) written by the AI, when you asked for one. Check its numbers against the results before you use it.
• A de-identification log, when the data were de-identified in Celeus, when the bundled copy of the data had identifiers removed for packaging, or when a flagged column was reviewed and kept.
• A transformation log, when the data were cleaned or transformed in Celeus.

Handle it like the data

The package includes a copy of your data, so treat it with the same care as the dataset itself. When the package is built, columns the identifier scan flagged and the analysis does not use are de-identified in that copy; your analysis input is not changed. A flagged column the analysis does use stays as it is, and the run’s flags say so. De-identify before you share a package outside your team.

Table 1 contains real aggregates from your data. Review it before sharing, too.

What the package is for

The package is what lets someone else check your work: a reviewer, a co-author or a statistician can see exactly what was run, on which data version, with which settings, seed and package versions. How that makes a result repeatable is explained in Determinism and reproducibility.