nature-statistics

Nature Statistics

Use this skill for statistical reporting and audit. Reanalyse raw data only when the user explicitly requests computation and supplies the design metadata needed for a reproducible analysis.

Core rules

• 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.

1. Extract groups, endpoints, hierarchy, repeated measures, blocks, missingness, exclusions, randomization, and blinding.
2. Define n and distinguish biological/experimental, technical, repeated, and nested observations.
3. Map each claim to its comparison/model, assumptions, correction, effect, uncertainty, and exact p-value policy.
4. Audit Methods, Results, legends, and source-data notes for consistency.
5. Draft conservative ready-to-paste text and list factual gaps as AUTHOR_INPUT_NEEDED.
6. 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.