Treat a reported number as the end of a chain
A finding emerges through several stages. Data is selected, cleaned and transformed; a procedure runs with particular options; software emits output; and a writer turns that output into a table, figure or sentence. A screenshot of the last screen hides all earlier decisions. Syntax on its own does not prove that the file generated the run ultimately reported.
Connect the artefacts with a model or result ID. For example, M03 appears in a manuscript note, syntax heading, output filename and analysis log. It points to a fixed data version and execution time. If an error is corrected later, the new run receives a new ID rather than overwriting the historical evidence.
Preserve raw output but design reader-facing results
Keep the machine output intact, including warnings, excluded cases, convergence messages and session information. It may be lengthy and unsuitable as a thesis table. Create a separate scholarly table that communicates the relevant quantities clearly and consistently. The raw output need not fill the appendix, but it should be locatable inside the reproducibility package where permissions allow.
Record the source of each transferred value: output object, table, row, coefficient name or stored estimate. Avoid typing by hand when a reproducible export can produce the table. Round only for display and calculate with unrounded values. If a reported figure is derived from several output values, preserve its formula and inputs.
Make model and case selection inspectable
Output documents one run, not why that run was selected. Maintain a model register containing the research question, outcome, predictors, interactions, weights and exclusion rules. Do not erase exploratory variants and sensitivity analyses. Distinguish planned analysis from investigation prompted by the observed data.
Log counts at each stage: imported, excluded, unavailable because of missing values and actually analysed. Procedures may handle missingness differently, so use the count reported for that model rather than copying the total number of source rows. Weighting, clustering and complex-survey options belong in syntax, method and the model register.
Use different software conventions for the same evidential goal
In SPSS, saved syntax should replace or record choices made through dialog boxes. The output file contains tables and messages but must be tied to the syntax actually run. Name sections clearly and make filter, weighting and Split File states visible. A stale filter can silently alter every later procedure.
In R, a script or reproducible document can connect import, transformation, modelling and export. Capture session and package information. A saved workspace image is not a readable method. In Stata, do-files and logs can record execution; establish clean starting conditions and avoid interactive edits that never enter the do-file. Across all three, syntax expresses the method, output proves the run and the report table communicates the selected evidence.
Check meaning as well as transcription
Inspect sign, reference category, scale, unit, decimal conventions and rounding. Check whether intervals or standard errors align with the estimate. Confirm that a p-value belongs to the named test and that multiple-testing procedures, weighting or robust standard errors are described correctly. Convergence, sparse-cell or singularity warnings must not disappear when output is reformatted.
Run the complete analysis into a fresh output directory so that only newly generated files can appear. Compare core objects or tables automatically where possible, or state numerical tolerances. Link the run to the software environment record. Stochastic procedures also need their generator and seed.
Write for readers without imitating the software window
A thesis table should follow disciplinary reporting expectations and include quantities required for interpretation. It does not need to resemble the output viewer. Explain variables and reference groups in notes, use consistent precision and place the result ID in a technical note or reproducibility index. Cite the software following its developer's preferred citation where this is customary in the field.
Keep interpretation separate from mechanical production. The audit trail can show that a coefficient was copied correctly and regenerated from the declared data. It cannot establish that the model answers the research question, assumptions are defensible or the effect is substantively important. Those arguments belong in methods, results and discussion.