Identify what was measured, the independent experimental unit, and the unit actually analysed. Do not silently promote cells, images, fields, readings, runs, or technical replicates into independent n.
Report effect size, uncertainty, sample size, test/model, assumptions, and multiplicity alongside significance where applicable.
Do not invent n, p values, confidence intervals, degrees of freedom, software versions, correction methods, exclusions, randomization, blinding, preregistration, or power calculations.
Do not make a final test recommendation when design or analysis unit is unclear.
Do not upgrade association to causality or significant to important, biologically meaningful, or large.

Route and workflow

Classify the task as audit, rewrite, draft, figure-statistics alignment, reviewer-response support, or data-backed reanalysis. Read manifest.yaml, load its minimum reporting sources, then only the condition-matched failure-mode, figure, or reviewer checklist.

Extract groups, endpoints, hierarchy, repeated measures, blocks, missingness, exclusions, randomization, and blinding.
Define n and distinguish biological/experimental, technical, repeated, and nested observations.
Map each claim to its comparison/model, assumptions, correction, effect, uncertainty, and exact p-value policy.
Audit Methods, Results, legends, and source-data notes for consistency.
Draft conservative ready-to-paste text and list factual gaps as AUTHOR_INPUT_NEEDED.
For reanalysis, treat inputs as read-only, record their hashes, and write code, environment, derived outputs, and the decision log to a new output directory. Do not overwrite source data, prior results, or manuscript files.

Output and completion

Return scope and boundary, design/unit readout, severity-ranked issues, ready-to-paste text, missing author facts, and residual reviewer risk. Finish when design, analysis unit, n, method, effect/uncertainty, multiplicity, and unresolved facts are explicit.

nature-statistics

Nature Statistics

Core rules