CSR number QC against the TLFs — every number traced to the table cell it came from
Every number in a report traced to the table cell it came from, so reviewers decide only the few that disagree.
See one case, screen by screen ↓A CSR section quotes hundreds of numbers, and every one of them has to be found in the TLFs: the right output, the right analysis set, the right arm and visit, the right run. A QC reviewer does it with the draft on one screen and the RTFs on the other, ticking numbers one by one, and typically spends two to five days on a single document.
The errors that get through are rarely typos. They are a responder rate copied from dry run 2 before a data query changed it, a per-protocol sentence quoting the ITT value, a safety percentage divided by the randomised N instead of the safety set, a confidence interval from the sensitivity analysis, a p-value printed as p=0.000. And when one table is re-run, nobody can say quickly where else its old cells are quoted — the synopsis, the congress abstract, the manuscript, the lay summary.
Estimated split for one CSR section, in working days, by hand and with the solution. The totals follow the working solution’s dashboard, which shows a median of 4.5 days by hand and 1.7 days with Results Check for a CSR section.
Agents find every number, match it to its table, check the house rules and the other documents, and write the QC report; a QC reviewer decides each finding and signs.
VLT-302, a psoriasis study of Veltrimab against placebo. Dr. Maya Chen’s Section 11 — Efficacy Evaluation, Draft v0.9, 38 pages, is due Oct 14. Here is how it gets through QC, screen by screen, in the working solution.
Lena opens Results Check: 1,150 numbers checked in the last 7 days across 9 documents and 3 studies, 18 findings waiting for her, one document due within 2 days. Each draft on the desk shows its verdict bar — CSR §11 Efficacy: 214 numbers, 205 agree, 9 to decide.
“Every number in your drafts, traced to the table it came from.”
The Number finder reads 214 numbers on 38 pages with their statistics, arms and visits. The Table matcher finds 213 of them in 16 outputs of the final run, and one with no output. The Rule checker applies 6 house rules and recalculates 5 derived numbers; the Cross-document checker compares the section with 6 other VLT-302 documents.
“9 findings ready for your decision, each with both locations.”
Page 47: sPGA 0/1 responders, Veltrimab, ITT set. A line runs from the sentence to the cell in Table 14.2.2.1. Six checks pass — output found, endpoint and visit, analysis, arm and column, population, rounding — and two fail: current run and value. The suggested fix is 66.6% → 66.4%.
“The draft quotes dry run 2 (Sep 12). In the final run (Oct 2) one Week 16 sPGA reading at site 114 was corrected after data query DQ-1187, so Veltrimab responders fell from 285 to 284.”
Lena switches the table from Final run to Dry run 2 · Sep 12. The superseded output opens stamped SUPERSEDED, with 285 (66.6) in the Veltrimab column — the cell the draft quotes, from data before database lock.
The dry-run table footer: “DRY RUN — NOT FOR REPORTING. Data before database lock.”
The draft says that in the per-protocol set 27.6% of Veltrimab patients reached PASI 100. That is the ITT value from Table 14.2.3.1. The per-protocol value in Table 14.2.3.2 is 116 (28.7). Population & N and value fail; everything else passes.
“Change the value — or the population and the table reference.”
When a finding needs a statistician’s call, Ask statistician drafts the query to Omar Haddad: the page, the value in the draft, the cell in the final run, and “Which value should the document quote?” The sentence and the table cell open highlighted for him.
Once the query is sent, Omar’s reply: “The sentence names the per-protocol set, so it is 28.7%. If Maya prefers the ITT value, change the population and cite 14.2.3.1 instead.”
“Response was durable: 88% of Week 16 PASI 75 responders maintained their response through Week 52.” The Table matcher searched all 412 outputs; the nearest, Table 14.2.1.3, stops at Week 16. The maintenance period is still blinded. With the eight suggested fixes for the other findings accepted in one step, Lena marks the sentence for removal — it goes on Dr. Chen’s corrections list.
“Week 52 tables are planned for the maintenance-period addendum. Until then the sentence has no source.”
The QC record can be signed only once every finding is accepted, kept with a reason or removed. 214 numbers checked, 9 findings decided, results package Final · Oct 2. Lena picks the meaning of her signature and re-enters her password.
Meaning of signature: “QC review complete — every number verified against the final tables.”
The QC report lists every finding with its page, the draft value, the table cell, the type and the decision — 641 → 642 count differs, 55.4, 67.0 → 55.9, 66.5 different analysis, p=0.000 → p<0.001 format, 88% sentence removed — then the numbers one by one with their source cell and verdict. It downloads as a PDF, with the comment log as .xlsx.
Table 14.2.2.1 changed in the final run: 285 (66.6) → 284 (66.4), and the difference 61.4 → 61.2. Its cells are quoted in the CSR, the synopsis, the congress abstract, the manuscript, the lay summary and IB §5.3. Re-check every quote re-runs all 9 against the final run and tells the writers.
“A data query at site 114 re-scored one sPGA reading. Every quote of the two changed cells is re-checked automatically.”
The same CSR section, seen by the five people who carry it — what their week looked like, and what it looks like now.
Agents find every number, match it to its table, check the house rules and the other documents, and write the QC report; a QC reviewer decides each finding and signs.
Extracts every number in a draft with its statistic, endpoint, arm, visit, population and sentence, and marks derived numbers with the cells they are built from.
Finds the output and the exact cell each number came from in the current results package, and records any superseded run that also contains the value.
Compares value, population, denominator, analysis, arm and visit with the house rules, recalculates every derived number, and gives a verdict, a reason and a suggested fix.
Checks that the synopsis, abstract, manuscript and summaries agree with each other and with signed documents.
Writes the number-by-number QC report, the comment log and the corrections memo for each writer.
Decides each finding and signs. The agents propose; a named person decides.
QC reviewers, writers and statisticians can ask in plain words — or tell it what to change. Every answer points to the sentence and the cell.
Where is Table 14.2.2.1 quoted, and what changed?
The final run (Oct 2) changed 2 cells: Veltrimab responders 285 (66.6) → 284 (66.4) and the difference 61.4 → 61.2 (data query DQ-1187, site 114). Those cells are quoted 9 times in 6 documents with the old values — CSR §11, the abstract, the manuscript, the lay summary, the synopsis and IB §5.3. The synopsis and IB corrections are already with the writers.
Why does the safety set matter?
Safety percentages use the safety set — the 427 Veltrimab patients who were treated — not the 428 randomised. Dividing by 428 gives 52.1% for any TEAE instead of 52.2%, and 12.1% for nasopharyngitis in the manuscript instead of 12.2%.
Add a rule: “about N times” must be within 5% of the table ratio
Done. New house rule: “about N times” must be within 5% of the ratio of the table values. CSR §11 says “about 10 times”, and 72.7 ÷ 7.0 = 10.4 — 3.8% away, so it stays derived · verified. The rule is live in Settings and on the audit trail.
Is CSR §11 ready for QC sign-off?
Not yet. 9 findings still need a decision. 8 have a suggested fix you can accept in one step; the Week 52 sentence has no source and needs removing or a reason.
13 screens from the working solution, on its sample data. Pick one to see it large.
What counts as a match, how numbers are rounded, and who decides are settings — each change recorded in the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Results package numbers are matched against | VLT-302 final run · Oct 2 | VLT-302 final run · Oct 2 · VLT-302 dry run 2 · Sep 12 |
| Flag numbers that match a superseded run | On | On · Off |
| Compare with signed documents | On | On · Off |
| Rounding | Half away from zero (SAS ROUND) | Half away from zero (SAS ROUND) · Half to even |
| Percentages in the body | One decimal (72.7%) | One decimal (72.7%) · Whole numbers (73%) |
| Small p-values | p<0.001 | p<0.001 · p<.001 |
| Treat format-only differences as notes | Off | On · Off |
| Who signs the QC record | Lena Ortiz | Lena Ortiz · Dana Okafor |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample study data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (sponsor results for AI-assisted report writing — about half the number errors reaching review; manual QC 2–5 days per document). People, companies and products named on this page are fictional — characters and sample data in the working solution.
It is the QC step that confirms every number quoted in a results document — a CSR section, synopsis, abstract, manuscript or lay summary — matches the tables, listings and figures it came from. Results Check does it number by number: it finds each number, locates its output and cell, and checks value, population, denominator, analysis and rounding.
The Number finder records each number with its statistic, endpoint, arm, visit and the analysis set named in the sentence. The Table matcher then finds the output whose title, population, endpoint and visit fit, and the row and column for the arm and statistic, in the current run. If nothing fits, it says which outputs it searched and shows the nearest.
Yes. Verdicts use the current run only, and a number that matches a superseded run is flagged as a stale table version, with the old run one toggle away. When a table is re-run, every quote of the changed cells, in every document, is re-checked.
By your house rules — for example, round once, half away from zero; one decimal for percentages in CSRs; p<0.001 for small p-values. Derived numbers such as differences, ratios and “about N times” are recalculated from the cells they are built from. Format-only differences can be treated as notes.
No. The agents only read drafts. Accepted fixes go to the document’s author on a corrections list, and the writer makes the change.
Yes. Only a named reviewer can accept a fix, keep a number with a reason, remove a sentence or send a query to the statistician, and only once every finding is decided can the QC record be signed — with an electronic signature that records the meaning and asks for the password again.
Any document that quotes results: CSR sections, synopses, interim reports, congress abstracts, journal manuscripts, plain-language summaries and Investigator’s Brochure sections, in Word or PDF, against TLF outputs in RTF or PDF.
The Agentic Solution Engine builds and deploys it from your requirements — your house style rules, QC report template and a results package with drafts that quote it — and it goes live once every quality gate has passed. We will walk you through it on one of your own drafts first.
A trial synopsis becomes a checked first-draft protocol in days, in your own approved language.
1-min filmTMF CheckEvery trial document is read, filed in the right place and checked, so the file is inspection-ready every day.
1-min filmCSR WriterResults sections and patient narratives are drafted within days of database lock, every number linked to its table.
1-min filmWe’ll run Results Check on one of your own drafts against its results package.