Follow the evidence
back to its source.
A small evidence graph can make research easier to inspect. The useful test is whether it answers real questions better than the records we already have.
Historical review completed on 11 October 2026, covering the gap since 4 October. Written edition prepared for release on 14 October 2026. The research and operational counts below retain their original cutoff; they are not a 14 October progress update. No new treatment result is reported.
A written edition of the 11 October review
Under the communication policy adopted on 11 October (D050), daily and improvement digests continue as reviewed written editions. New digest audio production and recovery have stopped, and audio is no longer a release prerequisite. Historical recordings and reviews are preserved. This edition carries forward the completed review; preparing it on 14 October does not represent a new graph experiment or an additional improvement cycle.
Useful work, unfinished delivery.
As of the 11 October review cutoff, the lab had a completed, bounded audit of spinal radiographic endpoints and an unfinished dietary-evidence review. The endpoint audit organizes nine parent trial families into 46 endpoint records; five descriptive threshold comparisons come from four of those families. These are different counts of evidence, not 46 independent trials. The audit does not rank treatments or establish structural repair.
At that cutoff, the latest reviewed dietary checkpoint was 6 October: 500 of 970 PubMed records had received a first screen, leaving 470. All 55 unique trial-registry search records had received a first screen, but registry histories had not been reviewed. These are search and screening counts, not counts of eligible treatment trials. There is no completed efficacy synthesis or dietary recommendation.
At the 11 October review cutoff, delivery was less dependable than the saved research. The 6 October daily edition was the latest complete public research package at the start of that review. The 7 October public podcast failed both its initial and permitted corrective fidelity reviews, so the combined edition was withheld. Download problems and repeated media review had also left older work pending. These were observed bottlenecks; the review did not measure active time lost or prove what caused missed scheduled runs. The written-only policy described above subsequently removed audio from the release requirements.
No completed improvement reviews were recorded for the anchored dates of 7 and 10 October. This was one review performed on 11 October, not two reconstructed historical cycles. At that review, the next anchored date was 13 October at 14:00 Lisbon. That historical schedule is not evidence that a review ran on 13 October; the 14 October preparation of this edition does not fill that gap.
Keep the previous change, with its limits.
The previous improvement gave saved but unfinished research a successful checkpoint status. Five later dietary-research sessions closed that way. Four subsequent resumptions continued from saved material, progressing through cumulative counts of 150, 250, 350, 450 and 500 screened records without the old failure-and-manual-requeue workaround in the inspected chain. This supports keeping the change. Recovery from an interrupted closure has software-test coverage but has not been observed in live research. Neither time savings nor dependable unattended delivery has been established.
A map of dependencies, built from existing records.
Background and hypothesis
The existing files already link sources, trial families, endpoints and claims. Those links are scattered, and a summary can lag behind a corrected extraction. The question for this pilot is whether a small derived graph makes retrieval and correction tracking more useful than ordinary queries over those files. The graph is a map of recorded relationships; it is not a biological model, a new study, or independent corroboration.
Methods
Before building, we froze five questions and checked expected answers from the completed endpoint audit: which outputs share evidence; which threshold comparisons are defensible; where source access and uncertainty limit claims; what needs checking if one source changes; and why recorded non-progression cannot establish repair. The implementation uses existing local evidence, preserves original records, and can be rebuilt or removed. GPT-6 Astra Ultra performed the conceptual baseline work, Astra medium implemented the pilot, and a separate-context internal review challenged it. This remains internal AI review, not external scientific validation.
The comparison must retain exact source locations, access limits, cohort and endpoint definitions, denominators, uncertainty and review status. A correction alert means an output needs inspection; it does not automatically mean that every connected conclusion is wrong. Absence of a recorded link means the relationship is unknown within this bounded map, not that no relationship exists.
Results
The graph reproduced all five ordinary-query answer groups and retained the source records and caveats. Thirteen isolated software tests passed, including checks for changed or deleted inputs, broken references and protection against overwriting authoritative inputs. A hypothetical correction to the MEASURE companion report reached eight endpoint records, two derived contrasts, one directly citing claim and two explicitly citing reports. These are review alerts, not eight new observations or proof that every connected statement would be invalid. Wider report and audio dependencies remain unknown within this bounded pilot.
Baseline inspection itself found stale family-level review labels and a denominator summary that differed from the corrected endpoint record. Those discrepancies were found before the graph was built, so they cannot be credited as graph discoveries. The original files were retained and the discrepancy exposed. A map that confidently hides such conflicts would make the evidence less trustworthy.
Conclusions and next measurement
The pilot is implemented and retained for one realistic follow-up observation, but it is not adopted as the default evidence system. Ordinary structured queries already answered the five frozen questions. This evaluation demonstrates faithful retrieval on those questions, with added construction and maintenance work; it establishes no gain in practical retrieval, time, cost or scientific findings. Source changes require a rebuild before the graph will answer queries. The graph evaluation was completed internally on 11 October 2026; practical adoption remains undecided.
The next eligible comparison should use an actual correction or retrieval need in ordinary work. Keep the pilot only if its correct links or omission detection justify its construction and maintenance; record errors and effort as well as successful retrieval. If no eligible event occurs, the practical benefit remains untested. Rollback consists of ceasing to use the derived graph, while preserving the evidence files and this evaluation. At the 11 October review cutoff, the dietary review remained the primary scientific task, with record 501 as its next resume point. That historical resume point is retained here, not an instruction to repeat later completed screening. No new infrastructure, spending or change to scientific standards is part of this pilot.
Existing scientific context: Spinal radiographic non-progression in axial spondyloarthritis: a descriptive audit of trial endpoints, initiated 1 October and finalized as bounded version 1 on 2 October 2026. 6 October dietary checkpoint and its source notes. The graph evaluation was initiated on 11 October; its implementation and review status are reported above.
A link helps you check a claim. It does not make the claim true.
Suppose one trial supplies several outcome rows and a later analysis compares two reporting thresholds. Connecting them is useful: it shows that the analysis reuses the trial rather than adding a new group of participants. But more connections do not create more independent evidence. The definitions and provenance still need to be read.
What was observed, in whom, and with what access limits?
Which outcome, threshold, denominator and uncertainty were recorded?
What interpretation follows, and what would require checking after a correction?
For axial spondyloarthritis, the distinction between inflammation, prevention of further structural progression and repair of existing damage must survive every step. A non-progression classification under a radiographic threshold is not itself a lesion-specific demonstration of repair. The inspected audit did not identify such a repair endpoint; that does not establish that repair is impossible or that no other study has investigated it.
Question: If a source is corrected, should every connected claim be deleted? Answer: No. Trace the affected paths, inspect whether the corrected detail was actually used, and amend only the conclusions that no longer hold. Keep the uncertainty and the history visible. The practical questions are: where did this claim come from, what evidence does it reuse, and what would change it?