# Write and report > Turn verified results into a shareable package, and check a manuscript’s numbers, tables, figures, and claims against their evidence—including after your AI rewrites it. ## Keep the manuscript aligned with the evidence. Compare text, tables and figures with their research evidence. See where numbers or wording drift, distinguish rule-based checks from AI assessments, and review corrections as a new document version. ## Know what to change, and why. Select a finding to trace the statement to its evidence and review the suggested correction. *Illustrative draft · synthetic study data · not a SciGround report.* **Three findings in this draft:** 1. **Sample count** — 12 → 10 animals per group (Rule-checked) 2. **Claim strength** — “Causes weight loss” goes beyond the study (AI-assessed) 3. **Error bars** — Needs the plot data to check (Not checkable) ### Draft > Day-14 body weight was compared between treated and control mice (n = 12 per group [1]). Mean body weight was 2.08 g lower in the treated group (95% CI 1.33 to 2.83 g; Welch’s t-test), showing that the treatment causes weight loss [2]. > > Figure 1. Day-14 body weight by group. Bars show mean ± error. [3] ### Compared with (Research version v3) - **Result:** Experimental unit: Animal; Animals: 10 treated · 10 control; Treatment − Control: −2.08 g; 95% CI: −2.83 to −1.33 g; Method: Welch, two-sided. - **Design:** Groups: Two independent groups; Outcome: One weight per animal, day 14; no baseline. - **Declarations:** Random assignment: Declared, not verified (Recorded as a declaration; the supplied files do not document it). - **Figure 1 file:** Contents: Image only (PNG); Plot data or script: Not supplied. ### 1. Sample count mismatch (Rule-checked) - **Why:** The draft says 12 per group; v3 has 10 animals per group, counted by animal. - **Suggested correction:** replace “n = 12 per group” with “n = 10 animals per group”. - **Compared with (v3):** Result: Animals, Experimental unit - **How it was checked:** A specified rule compared this item with its evidence. Sample-size rule: a count in the text must equal the cited result’s n for the same experimental unit. The rule compares this one count with the result; the rest of the study is checked by its own checks. ### 2. Claim goes beyond the result (AI-assessed) - **Each part of the claim, checked against the study:** - “2.08 g lower” — Result: **Supported.** Estimate and 95% CI (1.33 to 2.83 g lower) agree. - “weight loss” — Design: **Not measured.** One weighing per animal on day 14; no baseline, so no change within animals. - “causes” — Declarations: **Declared only.** A causal reading rests on random assignment, which is declared but not documented. - **Suggested correction:** delete “, showing that the treatment causes weight loss”, leaving “Mean body weight was 2.08 g lower in the treated group (95% CI 1.33 to 2.83 g; Welch’s t-test).” - With assignment records: “…, indicating that the treatment lowered day-14 body weight.” - **Compared with (v3):** Result: difference, 95% CI · Design: Outcome · Declarations: Random assignment - **How it was checked:** An AI assessment. Review its reasoning and suggestion; it can miss problems or raise false alarms. AI assessment of the wording against the result, design and declarations. “Weight loss” implies change within animals, but each animal was weighed once: the result is a day-14 difference between groups. “Causes” rests on random assignment, which is declared but not documented. A suggestion to review, not a proven correction. ### 3. Error bars cannot be checked yet (Not checkable) - **Why:** The caption does not define the bars, and Figure 1 was supplied as an image only. - **To check it, supply:** The plot data or script; The error-bar definition: SD, SE or 95% CI. Then the caption is checked against the plotted values. - **Compared with (v3):** Figure 1 file - **How it was checked:** The supplied evidence does not allow this check yet. The figure-definition check needs the plotted values or their source. Nothing supplied says whether the bars show SD, SE or a confidence interval, and the definition is not inferred from the image. The gap stays in the report until the material is supplied. Review corrections as a new document version, with the original preserved. Submit the revised draft to check it again. Findings cover the text, tables and figures you submit, and each says how it was reached: rule-checked, AI-assessed or not checkable. ## Map the draft to evidence. Review the correction. 1. **Submit — Send the draft and its evidence.** Text, tables, figures with source data, and the verified results. 2. **Compare — Map each statement to evidence.** Numbers, units, directions, and claims are matched to their sources. 3. **Correct — Review suggested corrections.** Corrections come as a new version with reasons and a diff. 4. **Recheck — Confirm the new version.** The corrected draft is checked again before it goes back to your document. ## A review you can act on - A shareable results package traced to evidence - A finding-by-finding report, each labeled rule-checked, AI-assessed, or not checkable, with its location and evidence - A corrected version and diff, with the original preserved - A section listing every pre-specified result, significant or not ## Document checks and correction delivery. ### A results package, ready to share. Generate a package with methods, results table, assumptions, limitations, checked scope, and the Record reference. Every number traces to its evidence, and display rounding never flips a decision. ### Manuscript checked against evidence. Submit a methods or results section with its tables and figures. SciGround checks numbers and rounding, sample sizes and experimental units, group direction, units, methods, axis and error-bar definitions, and how strongly each claim is worded. ### Catch AI drift every time you recheck. Rewritten and shortened text is checked again. Overstatement—non-significance as equivalence, association as causation, a limited result as a full guarantee—is flagged, and so is wording that weakens a sound result unnecessarily. ### Corrections, as a new version. SciGround drafts corrections for numbers, wording, and figure notes, keeps the original, and checks the new version again with a clear diff and reasons. ### Works where you write. Bring documents from GitHub, Word, or Google Docs with their versions. Corrections come back as a separate branch or pull request, a new document, or a diff—the original is never overwritten. ### Every planned result, on the record. The package lists every result you planned in advance, significant or not, drawn from the study’s own Records. A reviewer can fetch them by study ID and compare them with the manuscript. ## Limits and conditions Checking a manuscript against evidence is not the same as proving every sentence. See where the line is. [Read the limits](https://www.licklider.ai/sciground/writing/limits/) ([Markdown](https://www.licklider.ai/sciground/writing/limits.md)). ## Related - [Share and verify](https://www.licklider.ai/sciground/sharing/) - [Use with your AI](https://www.licklider.ai/sciground/connect/) - [Record](https://www.licklider.ai/record/)