The analysis is only half the deliverable.
A result that cannot be read, checked and submitted is not yet evidence. We author the document as carefully as we run the model — Word-native where the client works in Word, LaTeX or JATS where the publisher requires it, and reproducible either way.
Documents that survive review.
Each output has a governing reporting standard, and we write to it from the first draft rather than retrofitting compliance at submission.
Manuscripts & systematic reviews
Full IMRaD manuscripts, systematic review reports and evidence summaries, with claims calibrated to the design that produced them — no causal language on associational evidence, no effect sizes stripped of their intervals.
Detail
Reporting standards applied according to study design.
- PRISMA 2020systematic reviews and meta-analyses
- CONSORTrandomised trials
- STROBEobservational studies
- GRADEcertainty of evidence
- IMRaD structure
- Claim–evidence alignment
- Numerical consistency check
Protocols & statistical analysis plans
Pre-specified protocols and SAPs written before the data are opened, so the analysis that gets reported is the analysis that was planned. Deviations are recorded, not quietly absorbed.
Detail
Specified up front and carried through to the report.
- Estimandwhat quantity is being estimated
- Adjustment setderived from a declared causal structure
- Missing datamechanism assumed and handled explicitly
- Sensitivity analyseswhich assumptions are stress-tested
- Pre-specification
- Deviation log
- Reproducible pipeline
Policy briefs & executive reports
The same evidence, rewritten for a minister or a board without inflating what it supports. Uncertainty stays in the summary; it does not get lost between the appendix and the recommendation.
Detail
Written for decision-makers, traceable back to the analysis.
- Decision framingwhat choice the evidence informs
- Option appraisalcosted alternatives side by side
- Confidence statementhow much weight the finding bears
- Technical annexfull method for the reviewer
- Plain-language summary
- Costed options
- Traceable annex
Submission packages
Everything an editor needs, assembled and checked: cover letter, title page, blinded manuscript, tables, figures at the venue's dimensions, supplementary material, reporting checklist and declarations.
Detail
Checked before delivery, so the first editorial contact is a review and not a query.
- DOCXlive citation fields, not flattened text
- LaTeXwhere the venue requires it
- JATS XMLpublisher-ready structured export
- Vector figuressized to the venue, inspected at 100%
- Venue geometry
- Declarations & checklists
- Pre-delivery QC
Every reference is verified against a public record before it ships.
Fabricated and mis-attributed citations are the most common way a credible document fails. Each reference is resolved to a public identifier, checked against the metadata that identifier returns, and recorded in a ledger. Anything that cannot be resolved is reported as blocked rather than quietly included.
Identifier first, never title-matching alone
DOI, PMID, PMCID or arXiv identifier resolved through public metadata services. Title-and-year matching is used only to detect duplicates, never to assert a reference exists.
Metadata checked against the record it claims
Authors, year, journal, volume and pages are compared with what the identifier returns. Mismatches are corrected at source rather than in the bibliography, so the reference manager and the manuscript never diverge.
Retractions and expressions of concern surfaced
A retracted citation supporting a live recommendation is a serious defect. Retraction status is checked and reported, not assumed.
A ledger you can audit
Every reference carries its identifier, verification route and check date. Anything unresolved goes to a blocked manifest with what was attempted — the same discipline we apply to anomalous data points.
What gets stripped out.
No process narration
Tool names, file paths, pipeline commentary and internal QC labels never reach a reader-facing document unless an audit appendix was requested.
Numbers reconciled across the document
Abstract, results, tables and figures are checked against each other. Denominators and time points must agree everywhere they appear.
Claims calibrated to design
Observational evidence does not acquire causal verbs on the way to the abstract. Effect estimates keep their intervals, populations and comparators.
Analysis, figures, manuscript.
One practice from the question to the submitted package, so nothing is lost in the handover between them.