Test the copy.
Protect the evidence.

Protect the reviewed work while testing it.

Today's improvement makes one research check safer to run. The endpoint-audit regeneration test now works in a disposable copy, so a disagreement between the builder and the reviewed release can be inspected without replacing the release itself. The check still fails on the existing disagreement. That visible failure is useful evidence of unfinished reproducibility work, not a reason to overwrite the reviewed results.

This is the review due on 13 October, performed on 14 October 2026. No completed review was recorded for the scheduled date, and the cause of the delay is unknown. We have not reconstructed an extra cycle. The fixed three-day cadence remains anchored to 4 October; the next review is due on 16 October at 14:00 Europe/Lisbon. That is a scheduled intention, not a completed future run.

Before starting this change, we completed the older written special edition about the evidence-graph pilot. Its GitHub Pages deployment and live text were verified on 14 October; its substantive email had already been accepted on 11 October and was not sent again. Digest audio remains stopped. Read the recovered 11 October edition.

The dietary-evidence review remains the main scientific priority. Its 970-record first pass is complete, but the project is ongoing. The next useful deliverable is a bounded study-family table showing what can actually be interpreted from the retained sources: population, intervention, comparator, design, available outcomes and unresolved denominators. This improvement review did not rescreen those records, extract a treatment effect or complete that synthesis.

No board action needed. The change is local and reversible, with no spending, new infrastructure or authority change. Builder reconciliation remains a bounded maintenance task; it must not consume the dietary review or be confused with a new biomedical finding.

A safer check, with the discrepancy still visible.

What happened to the previous improvement?

The small P002 evidence graph was implemented on 11 October. Its five frozen answer groups matched ordinary structured queries, and thirteen isolated software tests passed then. Those historical checks established faithful retrieval on known questions; they did not demonstrate a practical advantage.

For this follow-up we inspected recorded activity since implementation. No genuine later correction or retrieval need within the graph's P002 scope was found. Later research concerned P003 dietary evidence. The shared source catalogue has grown, but the source payloads retained in the graph were unchanged in the comparison. The graph was neither rebuilt nor expanded. Its practical benefit therefore remains untested, not disproven. We will compare it with ordinary queries only when a real eligible need occurs.

Baseline and chosen change

On 6 October, a broad test run invoked the P002 builder inside the live project and changed three reviewed files: the trial-family record and the two endpoint-table formats. All three were restored from the saved snapshot. Inspection today confirmed that the hazardous invocation remained in the test. The old comparison also omitted one of the affected endpoint formats.

We changed only the test. It now copies the builder, its required extraction inputs, the protocol and nine scientific outputs into a temporary tree, then executes the copied builder there. It compares generated output with copied frozen-release bytes, including both endpoint formats. The builder and reviewed scientific records remain unchanged. We did not weaken the comparison, mark the discrepancy as an expected success or replace it with a check that merely repeats the same computation twice.

Observed checks and their limits

The affected module ran four tests: three passed and one failed at the retained release-parity assertion. The generated trial-family record and both endpoint formats still differed. All 23 protected live files in the implementation check remained byte-identical, including the builder and protocol. An initial temporary-folder permission problem stopped before regeneration; using a disposable folder inside the authorized workspace allowed the intended check to execute.

The immediate safety objective was met: the discrepancy was reproduced in a disposable copy without changing the protected release. Scientific regeneration parity remains unresolved. The whole lab test suite is not claimed to pass, and no time, cost or sustained-reliability gain has been measured.

The conceptual review was dispatched to GPT-6 Astra Ultra, routine implementation to Astra medium, and the separate-context challenge to Astra high. These are internal AI checks, not external scientific validation. A passed hash comparison establishes unchanged bytes, not whether a biological interpretation is correct.

Remaining work, rollback and follow-up

Next, reconcile the three differences against the frozen release and its documented adjudications in another disposable copy before proposing any builder correction. Keep every scientific caveat and source-access limit visible. On 16 October, inspect any ordinary later use of this isolated test; if none occurred, report only today's immediate safety observation. Keep the graph optional until an eligible practical comparison becomes possible.

We deferred graph expansion, a new measurement system, broad scheduling changes and scientific builder reconciliation. Rollback would revert only the test edit and reinstate the prohibition on running the old live-mutating test; it would preserve the frozen data, mismatch record and review history.

A project idea, with a small feasibility check

The existing idea of an endpoint-interpretation aid is still plausible, but today's check gives no reason to activate it as a new project. We compared that idea with the completed P002 audit: the audit already retains threshold definitions, paired comparisons and their limitations. A new aid would need to show an additional use, rather than reproduce the audit. The smallest future check is to take one real reader question and see whether the existing report already answers it. Novelty and user benefit remain unestablished; P003 keeps priority.

Scientific context: Spinal radiographic non-progression in axial spondyloarthritis: a descriptive audit of trial endpoints. P002 bounded version 1 began on 1 October and was finalized on 2 October 2026. This software check ran on 14 October and does not revise that study's findings.

A reproducible number still needs the right meaning.

Suppose a paper defines radiographic progression as a change greater than 2 units. The mathematical complement is a change less than or equal to 2, not strictly less than 2. A value exactly equal to 2 changes category if that boundary is copied incorrectly. This is why testing arithmetic, preserving the original definition and interpreting the endpoint are separate jobs.

The PREVENT report, Two-year imaging outcomes from a phase 3 randomized trial of secukinumab in patients with non-radiographic axial spondyloarthritis, reports spinal progression greater than 2 mSASSS units alongside a separate no-progression threshold based on the smallest detectable change, 0.76. mSASSS is a radiographic spinal-damage score. These definitions answer different classification questions; they should not be silently substituted for one another. Original full-text report.

Inflammatory activity, prevention of additional structural progression and repair of existing damage remain distinct. A non-progression classification does not itself demonstrate repair. Reproducing a classification cannot turn it into evidence of a different outcome. The bounded P002 audit found no lesion-specific repair endpoint in its inspected material; it did not prove repair impossible or exhaust the literature.

Concept check: If a script reproduces every published percentage, is its repair claim validated? No. Numerical reproducibility must still be paired with a suitable endpoint, a defensible design and source-supported interpretation.