# Stats Desk

> Statistical reporting for scientific manuscripts: a free in-browser recomputation of every
> reported p value and a reporting prescan, then a paid statistics audit or a drafted Statistical
> analysis section on gpt-terra.

URL: https://stats-desk.skillsafe.ai/
API tutorial: https://stats-desk.skillsafe.ai/api.html

## What it does

- Recomputation (free, in the browser, no AI): finds reported t(df), F(df1, df2), chi2(df), r(df)
  and z results with their p values and recomputes the p value each statistic implies - two-tailed
  for t, z and r, upper tail for F and chi-square - with the range the printed rounding allows. Each
  result is consistent, inconsistent, a decision error (the mismatch crosses 0.05), consistent only
  if one-tailed, an impossible value, or has no p. The distribution functions agree with R's pt, pf,
  pchisq and pnorm to a relative error below 1e-12 on 600 test cases.
  An adjusted p (Bonferroni-, Holm- or FDR-adjusted, P_adj, q) is only checked to be at or above the
  raw p the statistic implies. Degrees of freedom with thousands separators, rho/Spearman results,
  a semicolon before the p value and results wrapped across two lines are read. Anything that looks
  like a test result but cannot be recomputed (for example r = .41 with no df) is listed as
  "not checked" rather than silently skipped. Every line reference selects that line in the text,
  and the statistics CSV carries a what-to-do and an empty response column for a reply to reviewers.
- Reporting flags (ids S1.., priority P0 / P1 / P2): n counted as cells, spines, fields or
  technical replicates without averaging or a nested model (possible pseudoreplication); many p
  values with no multiple-comparison correction; p printed as zero; threshold-only p values;
  "ns" without a p value; trend language; a difference in significance read as a difference in
  effects; tails not stated; t or F without degrees of freedom; n as a range; no n at all;
  undefined error bars or plus-minus values; s.e.m.; normality asserted at tiny n; no software;
  vague exclusions; association written as causation; no randomization or blinding; undefined
  significance stars; and, for Nature, repeat counts for representative results.
- One-line checker: type "t(14) = 2.31, p = .036" and see the implied p.

## Paid lanes (field `task`)

- `audit` - scope (design and independent unit), issues ranked P0-P2 with verbatim evidence and
  line numbers, a claim-to-analysis map, a ready-to-paste revision with AUTHOR_INPUT_NEEDED
  placeholders, reviewer risk, and short author questions.
- `draft` - a conservative Statistical analysis subsection built from the text, the author's design
  notes and an earlier audit, with reporting notes and a figure-legend statistics template.

Every reply is reconciled in the browser: each flag answered, each confirmed P0/P1 flag carried by
an issue, the verdict (not_ready / needs_revision / ready) no looser than the flags left standing,
each quote present on its cited line, no number, test, correction or software in the ready-to-paste
text that the manuscript or design notes do not contain, and no p value in the review other than the
reported ones and the browser's recomputed values.

## Nature's statistical requirements (read on nature.com on 2026-09-27)

From https://www.nature.com/nature/for-authors/initial-submission ("Statistical information"): a
Methods statistics section describing the tests and whether they were one- or two-tailed; error
bars defined throughout the figures; EXACT n values for all statistics, individual values rather
than a range if n varied; the number of times representative measurements were repeated; exact
values for significant and non-significant P values where relevant; F values and degrees of freedom
for ANOVAs; t values and degrees of freedom for t-tests; replicates specifically defined. These are
applied only when the target is Nature.

## Sources

Derived from the agent skill @yuan1z0825/nature-statistics (yuan1z0825/nature-skills, Apache-2.0):
- https://skillsafe.ai/skill/@yuan1z0825/nature-statistics
- https://github.com/yuan1z0825/nature-skills