Who entered each treatment group?
The assignment count does not establish every analysis count.
20 October 2026 · Operational improvement special edition
Keep the explanation tied to the evidence
Today’s improvement review resolved how a disputed denominator caveat should be described. The proposed wording remains separate from the historical research release; the rebuild discrepancy is still open.
CEO report
The lab needs to reconstruct research tables without changing their meaning. The previous review located a small set of disagreements between an endpoint table released earlier and the output of its current builder. Matching numerical values did not settle whether the surrounding explanations were accurate. Repeating that build would add little, so today we checked the most consequential disputed explanation against the original paper.
This is one current review for the missed 19 October date, performed on 20 October. The every-three-day cadence stays anchored to 4 October; the next date is 22 October at 14:00 Europe/Lisbon. The earlier noon input checks and implementation of stronger scientific-review controls were separate work. They did not discharge this review.
The selected improvement is a versioned source-reconciliation record. It preserves every disputed field, records which wording the original source supports, and identifies what remains unknown. This reduces one documentary uncertainty. It has not established a repaired pipeline, general reliability gains, time savings or external scientific validation. No new decision or routine approval is needed from Pawel.
Scientific work retains priority: genomics next has one limited descriptive extraction under its current protocol. The history project remains stopped for the previously requested source. This review changes neither research scope nor writing permission. Today’s first observed noon action checked input readiness; future timely unattended operation remains unproved.
What we checked and what changed
On 17 October, one isolated build found 19 family-level text differences across three output files. Twelve of those differences also appear in two endpoint-table formats, so repeated representations are not extra findings. The numerical endpoint definitions and values inspected then were unchanged. Today reused that saved diagnosis and did not run another build.
The central source question concerned SURPASS, a trial comparing secukinumab with an adalimumab biosimilar using spinal radiographic progression. We inspected its original main article, focusing on the Statistical analysis section, Table 2 footnote and a separately reported sensitivity analysis. The footnote identifies the displayed counts as baseline measurements; it does not, by itself, establish the exact denominator used for every estimated percentage. The separate sensitivity passage needs a separate caveat because a Table 2 explanation does not document that analysis. Read the original SURPASS article.
| Disputed fields | Today’s decision |
|---|---|
| One family-level denominator note | Propose wording that keeps baseline measurements distinct from an exact outcome-analysis denominator. |
| Nine Table 2-derived endpoint reasons | Retain source-specific caution: preserve baseline counts separately and leave the exact endpoint denominator unknown. |
| Three sensitivity endpoint reasons | Propose wording tied to their Results passage; do not infer their denominator or estimator from Table 2. |
| Six corrupted-text fields | Unresolved. No automatic punctuation repair or source claim without inspection of the relevant originals. |
These are 13 proposed source-backed text decisions and six explicitly unresolved fields, not 13 corrections applied to the research outputs. All exact numerical denominators in the disputed endpoints remain unknown. The preserved release, extraction inputs, builder and failing comparison assertion stay intact. A separately reviewed correction or versioned successor is still required before the live mismatch can be called repaired.
The record was frozen before today’s new source inspection, but after the historical mismatch and values were already visible. It is a prospective plan for this documentary review and a retrospective assessment of earlier work, not prospective registration of the original study. The prior isolation safeguard remains useful for protecting files; no ordinary intervening parity-test use was found in the scanned canonical closure records. That absence means its ordinary-use benefit remains unmeasured. The optional evidence graph was not rebuilt or queried.
Next, use this decision record when a finite, separately reviewed repair is selected. Recheck the same preserved comparison against the original reference; an amended version must explain the distinction. At the next anchored review, assess actual use and upkeep. If no opportunity occurred, report it as unmeasured. Rollback simply stops using the derived proposal while retaining the original evidence and attempt history.
Why a percentage needs more than a nearby participant count
A structural-progression study compares spinal images taken at different times. Before interpreting a percentage, we need to know which people and which analysis produced it. The number assigned to a treatment, the number measured at baseline and the population used to estimate an outcome can differ.
In SURPASS, the Table 2 footnote explains that its displayed n counts patients with baseline measurements. The main analysis also describes multiple imputation: a method that represents missing values through several plausible completed datasets and combines the results. An adjusted or imputed percentage should therefore not automatically be treated as an observed integer numerator divided by that nearby baseline count. The separately reported sensitivity result needs its own denominator and analysis description. SURPASS original article: Statistical analysis, Table 2 and Results.
Schematic: different counts answer different questionsThese boxes show conceptual distinctions, not measured participant flow or equal populations.
Who had the starting measurement?
The Table 2 footnote explains this displayed n.
Missing values may be imputed. An adjusted percentage need not be a directly observed integer fraction.
This is a separately reported analysis. Its denominator needs its own evidence; do not borrow the baseline count.
Unknown stays unknownuntil estimate-specific evidence is available.
This matters because changing a radiographic threshold or the handling of missing images can change who is represented and what the reported percentage means. Two percentages from the same treatment group and follow-up time are not automatically comparable as nested counts. Keep the exact outcome definition, analysis population and missing-data method together. A result about limited new radiographic change also does not, by itself, show reversal of existing structural damage. The present lesson explains interpretation; it reports no new treatment comparison or repair finding.