Making a research correction usable
We now have a reproducible, explicitly labelled preview that applies the reviewed SURPASS wording corrections while preserving every numerical result. The original research release and its unresolved regeneration mismatch remain intact. This is a bounded software-and-provenance improvement, not a new biological finding.
Why this was the useful next improvement
The lab has two research lines. Genetics is waiting for specific original material from the Yang study; the supplied images helped but did not resolve that source gate. The history chat is independently developing its illustrated narrative. Neither needs another management feature today. A useful maintenance task was already waiting: make a previously reviewed correction usable without overwriting the evidence from which it arose.
The problem began when a script that regenerates the saved files from our earlier audit of clinical-trial outcomes produced wording different from the saved release. An earlier improvement isolated that test so it could fail without replacing the release. The next review located the differences. On 20 October, we checked the SURPASS paper and proposed thirteen wording changes, but applied none. A proposal alone did not provide a corrected dataset that another researcher could inspect and reproduce.
Today we selected one finite step: apply those thirteen decisions to a separate, versioned preview. We deferred rewriting the original builder, replacing the historical release, changing six unresolved text-encoding fields, and expanding the evidence graph. The expectation is a correction that can be examined and replayed. Time savings and ordinary-use benefit have not been measured.
What changed, and what it means
The scientific distinction behind the correction
SURPASS compared treatments while assessing changes in spinal radiographs. Its Table 2 lists how many patients had particular measurements at baseline. That count is useful context, but it does not by itself identify the exact patient group represented by every estimated result. The methods distinguish the full analysis set—the population used for specified analyses—from the safety population, and describe imputation, a method for filling in missing observations, for analyses in the testing hierarchy. We must preserve those distinctions instead of turning a nearby count into an unsupported denominator.
All twelve SURPASS denominator fields therefore remain unknown. The recorded baseline imaging counts stay 283, 280 and 283 for the three arms. The revised notes explain why those counts have not been substituted. They also distinguish three sensitivity-analysis percentages reported in the Results paragraph from the main table; that paragraph does not provide their numerical denominators. These are source-provenance clarifications, not new treatment comparisons. See the SURPASS primary report, DOI 10.1002/art.42852; this review used the retained publisher-text captures inspected on 20 October, including Methods and Table 2.
A correction that can be replayed
The preview contains the same nine study families and 46 endpoints as the original files. It changes thirteen wording fields in the family file and their twelve repeated representations in each of the flat JSON and CSV files: 37 representations of thirteen decisions, not 37 independent findings. Complete comparisons checked field identities, order, types, numerical values and all non-target text. Two independent output destinations produced byte-identical files. The six unresolved encoding fields remain unchanged.
| Check | Observed outcome | What remains unproved |
|---|---|---|
| Source and semantic review | A separate-context reviewer inspected the retained publisher passages and independently compared the original and preview files. | Internal review is not external scientific validation. |
| Replay and protected inputs | The same frozen inputs reproduced the same preview; the original release and builder were preserved. | The original builder still has an unresolved same-version mismatch. |
| Deliberate faults | Duplicate identity, invented denominator, changed rate, wrong sensitivity provenance and altered patch reference were rejected. | These selected controls do not measure the rate of errors in ordinary lab work. |
| Failure, repair, same-check repeat | The first checker accepted a false completion flag. That failure and the original checker were retained. After a metadata check was added, the same false flag was rejected and the valid preview still passed. | This is one demonstrated repair, not proof that every misleading statement will be detected. |
The original criteria and prior source decisions were retained. The additional metadata check was recorded after the reviewer exposed its omission; it is not presented as a prospectively specified success. Prototype work was prepared outside the shared lab records while the history chat was active. Its actual timestamps are retained separately from the later central ownership and import.
Did the result answer today's need?
Yes, within the selected scope: the lab has moved from proposed wording to an inspectable, reproducible correction preview. It has not reconciled the entire historical builder or completed the wider repair task. The previous interval contained no recorded ordinary use of the comparison between saved and regenerated files or an applied correction; an absence from the inspected records does not prove that unrecorded use never occurred. Today's deliberate checks are not counted as ordinary-use benefit.
The next anchored review is 25 October. It should look for genuine downstream use, discover whether the preview reduced ambiguity, and decide whether a complete versioned successor is worth adopting. Until then, the preview must stay clearly separated from the canonical research release. Rollback means ceasing to use this derived copy while retaining all originals, source decisions and failed checks. No founder decision is needed for this bounded change.
A short research lesson: a count is not automatically a denominator
In an experiment, the number of objects available at the start, the number actually measured, and the population represented by a statistical estimate can differ. Clinical trials add reasons such as missing scans and prespecified analysis populations. A percentage with a model-based estimate cannot safely inherit whichever patient count appears nearby. Leaving an unsupported quantity unknown preserves the question for later evidence; filling it with a plausible count hides the gap.
This matters for SpA research because apparent changes in spinal structure and measures of active inflammation answer different questions. Today's correction adds neither a treatment ranking nor evidence of structural repair. It makes the provenance of an existing result more explicit so later interpretation starts from what the source actually supports.