When the numbers agree but the explanation changes

Find the cause before changing the research.

The lab needs to be able to recreate a reviewed research table without losing the qualifications that make its numbers interpretable. Our earlier audit of spinal X-ray outcomes had a known problem: rebuilding it produced three files that differed from the saved release. The previous improvement made the test safe by running it in a disposable copy. Today we asked the next question: what, exactly, changes?

The diagnosis narrowed the problem. The numerical results and endpoint definitions in the endpoint tables were unchanged, but explanatory text was not: several strings contained different forms of garbled punctuation, and descriptions of why an analysis denominator was unknown differed. This gives us a concrete repair target. It does not yet make the builder reproduce the reviewed release, and we have not replaced either the builder or the scientific records.

The main research direction remains the existing plan to map established SpA manifestations and the sources that describe them before looking for overlooked symptoms. The bounded dietary inventory and Catalog annotation note remain saved; their broader research questions remain unfinished and deferred. This maintenance review neither restarts those branches nor starts the manifestation map.

No board action is needed. One limited diagnostic measurement was enough for this session. The next maintenance step is a reviewed correction in a disposable copy, with exact preservation of scientific caveats. It should not expand into a new data system or displace the selected research task.

From three differing files to an explained discrepancy.

What the previous change actually achieved

On 14 October, the regeneration test was changed to copy its inputs and execute the builder in a temporary working tree. That implementation check preserved all 23 protected live files while still exposing the existing mismatch. We found no recorded ordinary later use of that particular test before this review. Its longer-term practical benefit therefore remains unmeasured; today's separate diagnostic run is evidence about this session, not proof of routine reliability.

One measurement, with the original evidence protected

Before running anything, we saved the hashes that identify the exact builder, inputs, protocol and reviewed outputs. We then ran the copied builder once and compared nine scientific output files. Matching records by their trial-family and endpoint identities let us distinguish a changed explanation from a missing observation or a changed number.

Six outputs were byte-identical, including both formats of the five paired threshold comparisons. The same three files still differed: the trial-family record and the endpoint table in JSON and CSV formats. All 46 endpoint identities were preserved. Checked rates, means, counts, uncertainty fields, time points, thresholds, populations and analysis definitions were unchanged.

The family record contained 19 differing text fields: six strings with differing garbled punctuation in image-window, source-location and supplement prose, one SURPASS family-level denominator note, and 12 SURPASS endpoint explanations for an unknown denominator. The same 12 endpoint explanations also occur in the two endpoint-table formats; these are repeated representations of the discrepancy, not 24 additional scientific findings. No fields or source-access labels disappeared. That does not make the text changes harmless: the frozen SURPASS family note implies a denominator that its numerical fields leave unknown, while three rebuilt sensitivity-row explanations introduce a generic table reference where the endpoints cite a results paragraph. Both sides require source-specific reconciliation.

The differing text already exists in the saved extraction inputs. The builder passes it through. For SURPASS, its special denominator correction runs only when a numerical denominator is present; the current input denominators are already empty, so that branch does not normalize the accompanying explanations. The saved adjudications—review decisions about ambiguous or conflicting evidence—explain the scientific caution, but are recorded by the builder as an input hash rather than applied as a general set of correction instructions. Today's inspection therefore localizes the mismatch; it does not establish the exact historical editing sequence.

The observed benefit is a specific, reproducible diagnosis in place of a generic mismatch. All 23 protected live files and the isolated test remained unchanged. Regeneration parity is still unresolved. No time saving, cost saving or general reliability improvement was measured.

The evidence graph met a real request—and could not answer

Explaining the discrepancy required finding the saved adjudications and caveats, so this session provided a genuine retrieval need for the small evidence-graph pilot, which stores links among sources, records and conclusions. We attempted its existing verified query once. It stopped before returning an answer because the shared source catalogue no longer matched its recorded hash. A read-only comparison found 33 added source records while all 50 source payloads retained by the pilot were unchanged.

The direct route—reading the five saved adjudication records—remained available. The graph was not rebuilt, expanded or queried around its safety check. This attempt demonstrates a maintenance obstacle in this particular use, not a general verdict against graphs or proof that the current graph could safely ignore the change. It remains optional, with no retrieval advantage established. Its narrower hash design is a deferred alternative, not a second change implemented today.

Checks, rollback and next observation

The first comparison helper stopped after the successful copied build because two comparison rows share a trial-family identifier. We corrected the helper to use record identifiers only where they are unique, then analysed the saved output without rebuilding. A separate-context reviewer checked the diagnosis and the written explanation. The conceptual review used GPT-6 Astra Ultra, implementation used Astra medium, and challenge used Astra high. These are internal AI checks, not external scientific validation.

We deferred builder repair, scientific re-adjudication, graph redesign and scheduling changes. The diagnostic files are derived artifacts: rollback means ceasing to use them while retaining the evidence and audit history. The safe isolated test stays in place. A future correction should reproduce the frozen release with the reasons for missing denominators intact; if a scientific change is required, it needs a separately reviewed version rather than silently changing the old result.

This review was actually performed on 17 October, covering the missed 16 October anchor once. The reason for the missed execution is unknown. The 13 October review was already covered on 14 October. The cadence remains anchored to 4 October, every three days at 14:00 Europe/Lisbon; the next date is 19 October. A scheduled date does not prove future execution.

A possible next use, without another project

The existing idea of an endpoint-interpretation aid overlaps with the completed audit. Today's small feasibility check asked whether its numerical examples could be reused unchanged: the five saved comparisons matched, but the raw regenerated explanatory text did not. This supports using the reviewed release as the reference and postponing an automated aid. Its smallest useful future test remains one actual reader question that the current report cannot answer. No new project or claim of novelty follows from this diagnosis.

A denominator is part of the claim.

A percentage needs both a definition and a population. In the SELECT-AXIS 1 two-year report, 89.7% met a spinal X-ray score-change threshold below 2, while 76.5% met the stricter threshold of no increase. The modified Stoke Ankylosing Spondylitis Spine Score (mSASSS) summarizes radiographic damage. Table 1 reports 136 people with paired images for the mean-change analysis; that number is not separately restated beside the two percentages. Our audit therefore retains the linkage as an explicit inference. Original full-text report, spinal-radiograph results and Table 1.

The 13.2 percentage-point difference illustrates how a threshold changes classification. It is not a comparison of two treatments, an inflammation result or evidence that established damage repaired. This is why preserving a caveat matters even when every number survives a rebuild. When reading a percentage, ask: who was counted, under which definition, and is that denominator stated directly or inferred? Reviewed P002 interpretation and limitations.

Source: van der Heijde et al., Upadacitinib in active ankylosing spondylitis: results of the 2-year, double-blind, placebo-controlled SELECT-AXIS 1 study and open-label extension (2022), DOI 10.1136/rmdopen-2022-002280. Relevant original full-text passages rechecked on 17 October 2026. This lesson teaches interpretation; it gives no medical-care recommendation.