Methods and results: what you can rephrase and what must stay exact

Methods and Results are not ordinary explanatory prose. Methods define what was planned and done; Results report what the analysis produced. A stylistic edit is acceptable only when a competent reader would reconstruct the same sequence, materials, population, variables, analysis and findings from both versions.

Before revision, identify the source of truth for each type of statement. The protocol or preregistration governs prespecified design and outcomes; analysis code and a dated analysis plan govern computations; lab records or operational records govern what occurred; locked tables and figures govern reported values. The existing sentence is not automatically authoritative if it already conflicts with those records.

Do not use a lower detector score, smoother sentence or shorter word count as evidence that a scientific edit is correct. Correctness is tested against the study record.

Create an invariant register before editing: exact sample counts, group labels, variable names, units, time points, threshold definitions, effect measures, confidence intervals, p values, model names, exclusions and procedural sequences. The register is not a replacement for reading; it is a high-risk queue that makes omissions and altered tokens easier to find in Word Compare.

A reporting guideline tells authors what information a complete report should contain; it is not a paraphrasing template. The EQUATOR Network library is explicitly designed for transparent health-research reporting and can be browsed by study type, clinical area or report section. Use it only when its domain and study design match the manuscript.

CONSORT 2025, for example, asks randomized-trial reports to distinguish prespecified from post hoc analyses and to report each outcome's measurement variable, analysis metric, aggregation method and time point. STROBE asks observational reports to define outcomes, exposures, predictors, confounders and effect modifiers. PRISMA asks systematic reviews to present full search strategies and explain study exclusions. Each phrase carries study-design information that a generic rewrite can mistakenly compress.

These are health-research examples, not universal scientific law. A chemistry experiment, economics model, qualitative interview study or engineering benchmark has different reporting objects and may use a discipline standard, repository checklist or journal author guide. The transferable principle is narrower: identify the applicable completeness contract before changing prose.

A checklist can reveal an omission; it cannot certify that the reported method was actually performed or that the data are correct. Verify those against records.

Classify text by the consequence of a small wording change, not by section heading alone. A Results transition can be low risk; one word inside the same paragraph can reverse the effect. Use the table as a routing sheet and move any uncertain item to the stricter column.

“Freeze” means preserve the content byte-for-byte in the revision scope, not hide it from the final paper. A frozen equation or table can remain surrounded by revised explanatory prose. A high-risk paragraph can also be frozen as a unit when every sentence depends on exact thresholds, steps or values.

Do not infer that prose without digits is low risk. `Participants with prior exposure were excluded` encodes an eligibility rule; `analysts were blinded` encodes who lacked what information; `secondary analysis` encodes status. Conversely, a sentence containing a stable year in a background transition may be relatively easy to verify.

Read the revised Methods as if you had to reproduce the work without access to the author. Every procedure needs an actor or object, order, condition and enough operational detail for the applicable field. If reordering sentences makes chronology ambiguous, restore explicit sequence markers rather than assuming the reader will infer the original order.

• Population and sampling: preserve source population, setting, dates, recruitment route, inclusion/exclusion criteria, matching or sampling scheme and final analysis set.
• Materials and intervention: preserve manufacturer/model when required, software/version, reagent identity, dose/concentration, route, duration, frequency and comparator details.
• Variables and outcomes: preserve operational definitions, primary/secondary status, measurement instrument, metric, aggregation and time point.
• Bias controls: preserve randomization method, allocation concealment, blinding target, calibration, counterbalancing or other design-specific controls.
• Analysis: preserve model/test, covariates and why chosen, transformations, category boundaries, missing-data method, correction, sensitivity/subgroup plan and prespecified/post hoc status.
• Deviations: preserve what differed from protocol and why; do not rewrite an actual deviation into the planned procedure.

Compare verbs as carefully as numbers. `Randomly assigned` makes a methodological claim that `allocated` does not; `excluded before analysis` differs from `lost to follow-up`; `adjusted for age` differs from `controlled demographic factors`. A more general verb often deletes the exact decision boundary readers need.

For systematic reviews, treat database search strings, dates, filters and limits as executable methods. PRISMA 2020 asks for all information sources and the date each was last searched, plus full strategies. Do not paraphrase Boolean operators, field tags or database syntax. Explain them around a frozen search block if needed.

Cross-check Methods against the protocol, registration and analysis plan, then cross-check it against Results. Every reported subgroup or sensitivity analysis should have its status visible; every primary outcome should resolve to a method. If the documents disagree, do not edit for apparent consistency - document and resolve the discrepancy.

Trace each result as a tuple: analysis population + outcome/time point + estimate + unit + denominator + precision/uncertainty + inferential status. A revision passes only when every applicable element remains the same and the surrounding verb does not imply a stronger conclusion.

Do not treat a p value as a free-standing result. Keep the estimate and its precision where the field and checklist require them, and retain the actual comparison and analysis population. CONSORT 2025 asks randomized trials to report, by group, the number analysed, available outcome data, result, effect estimate and precision. STROBE asks observational studies to distinguish unadjusted and adjusted estimates and identify adjusted confounders.

For a systematic review, the result is not only a pooled number. PRISMA 2020 also covers the search and selection flow, excluded near-eligible studies with reasons, characteristics and risk of bias of included studies, heterogeneity, sensitivity analyses, reporting-bias assessment and certainty. A shorter rewrite that keeps only the pooled effect can materially misrepresent the evidence.

If the revised prose conflicts with a locked table, do not choose whichever looks more plausible. Return to analysis output and reconcile every occurrence.

The examples below are deliberately simple. Each revised sentence is grammatically acceptable; the failure is scientific. Use them as pattern detectors when reviewing your own comparison file.

Other high-risk swaps include `median` to `average`, `odds` to `risk`, `adjusted` to `corrected`, `participants` to `patients`, and `no evidence of a difference` to `evidence of no difference`. These pairs are not interchangeable. Search comparison markup for changed statistical and design terms even when no digit changed.

A rewrite can also preserve every sentence but move one into a new section. Moving an exploratory result into the primary-outcome paragraph or a limitation into Methods changes its rhetorical status. Review headings, table placement and paragraph order along with wording.

Keep the original and revised DOCX, then generate a third document with Word Compare. Review large deletions first, then run the invariant register through the markup. For every changed numerical or methodological statement, point to the protocol, record, code, output or table that verifies it. Do not approve from memory.

• Methods pass: another specialist would reproduce the same operation and decision rules from the revised text.
• Results pass: every value, denominator, unit, direction and uncertainty term reconciles with output and all document objects.
• Checklist pass: no required design-specific item was compressed away or reassigned to the wrong section.
• Language pass: causal strength, qualifiers, negative findings, exploratory status and limitations remain equivalent.
• Document pass: tables, figures, captions, footnotes, fields and supplements still resolve and agree.

Freeze the whole paragraph when it contains a dense protocol definition, executable search strategy, a chain of ordered steps, numerous interdependent values, or wording already aligned line-by-line to a registration or reporting checklist. Also freeze it when the original and underlying record cannot be reconciled: revision should wait for an author or analyst to resolve the discrepancy.

The button below opens the research-report humanizer. It is for prose that genuinely needs revision, not for changing locked data or manufacturing a preferred detector outcome. Exclude frozen material from revision where the workflow permits, compare the returned document, and accept only changes that pass the source-of-truth checks above.

The safest Methods or Results paragraph is sometimes the unchanged one. Revision has value only when clarity improves without changing what was done, observed or justified.