RevMan now asks why you downgraded

RevMan 11.8.0 shipped on 25 August. Most of the release reads as steady progress: dynamic Summary of Findings tables now work in every review type except diagnostic test accuracy reviews and overviews, and data packages containing dynamic SoF tables can be imported. Administrators gain a "Revert for amendment" option in version history. The preceding release, 11.7.0 on 19 August, was infrastructure only.

The change with real consequences for work in progress is smaller and easier to miss. RevMan now raises a validation warning when a certainty-of-evidence domain carries no documented rationale.

If you have a Cochrane review in flight, you will meet this warning.

Why an undocumented downgrade is a problem

GRADE assessments fail peer review in a predictable way. The certainty rating is there, the domain is marked as downgraded, and the reason is either absent or given as a single word: "imprecision". A reader cannot check that. They cannot tell whether imprecision was judged against an optimal information size, a confidence interval crossing a clinical decision threshold, or a small pooled sample that made the assessor uneasy. Those are three different judgements with three different implications for the reader.

Moving that check into the software changes when the problem is discovered. Until now it surfaced at peer review, at the point where fixing it means reconstructing reasoning from months earlier, often after the person who made the call has moved on. From 11.8.0, it surfaces while the review is open.

A rationale that will survive scrutiny records four things: which domain was downgraded, what specific evidence triggered the judgement, how many levels and why that number rather than one more or one fewer, and whether the same standard was applied to the other outcomes in the table. The fourth is the one most often skipped, and inconsistency across outcomes within a single SoF table is a reviewer's easiest catch.

The other line in the release notes

Cochrane also updated the copyright statements attached to exported files, which now prohibit AI and text-and-data-mining use.

This did not come with an announcement, and it will not affect most authors' day-to-day work. It matters for anyone building tooling on exported RevMan data, and it is a useful signal of where a major evidence producer has landed on the question of whether its outputs are training material. Expect other producers to follow.

Nested Knowledge changed its models under the hood

Releases 1.115.0 and 1.115.1, on 23 and 24 August, add Manual Usage Rights and Supplemental Materials support in ASTs. The substantive change is elsewhere: a model upgrade across the whole AI stack, covering Smart Screener, ASTs, Smart Critical Appraisal, Smart Insights and Bibliomining. The release describes the new models as roughly 50% faster, with improved instruction-following and modest gains in recall.

Faster and slightly better is welcome. The implication is less comfortable. If you benchmarked Nested Knowledge screening before 23 August and used that benchmark to justify a screening approach in a protocol, the benchmark now describes software that has been replaced.

This is a structural problem with AI-assisted synthesis rather than a complaint about one vendor. Platform models are upgraded silently and often. Any pilot study that establishes recall on a labelled subset has a shelf life measured in weeks, and the honest way to report one is with the date and platform version attached. Re-baselining after an upgrade is not paranoia; it is the only way the pilot means anything.

A new tool worth knowing about

TrialScout, published in the Journal of Clinical Epidemiology on 26 August by von Schreeb, Bruckner, Aspromonti, Caquelin, Cummins, DeVito, Axfors, Ioannidis and Nilsonne, is an LLM tool that matches published trial reports to their registry entries.

This addresses a bottleneck that anyone who has attempted an outcome-reporting-bias assessment will recognise. The methodological work of comparing pre-specified outcomes against reported ones is straightforward; finding the registration in the first place is what consumes the afternoon. Trial reports cite registrations inconsistently, registration numbers are mistyped, and some trials are registered on platforms that do not surface well in search.

The tool is open access and the code is on GitHub at lahnstrom/trialscout. For teams doing ROB-ME assessments or any work that depends on locating registrations at scale, it is worth an evaluation.

Quiet elsewhere, and one gap in what can be verified

The R ecosystem published nothing new in the window. Current CRAN publication dates: metafor 5.0-1 (26 April 2026), meta 8.5-0 (25 May 2026), netmeta 3.6-1 (20 June 2026), robvis 0.3.1 (2 June 2026), metaumbrella 1.1.0 (28 February 2025). dmetar remains available from GitHub only.

Also checked with no new release: ASReview (3.0.8, 18 June 2026), JASP (0.98.1, 7 July 2026, with the last meta-analysis change back in 0.96.0 in March), Laser AI (15 June 2026), DistillerSR (agentic AI launch, September 2025), Stata and Silvi.ai. Rayyan's last product update remains August 2024. Elicit posted a comparison of its own retrieval against web-search APIs on 14 August, reporting 60.3% against 40.3% recall on BioASQ; it is vendor-published and thin.

Two platforms could not be verified either way. Covidence and EPPI-Reviewer both serve release information on undated or JavaScript-rendered pages. That means no dated evidence of a release was found, which is not the same as confirming there wasn't one, and it would be dishonest to list them as quiet.

That gap is worth naming for a reason beyond bookkeeping. A meaningful share of the tooling that evidence synthesis now depends on does not publish dated, machine-readable release notes. Cochrane does. Nested Knowledge does. For much of the rest, a team trying to record which version of which tool produced a screening decision has no reliable source to cite. Reproducibility in AI-assisted synthesis will require vendors to publish versioned, dated changelogs, and most of them currently do not.

References

Cochrane. (2026, August 25). RevMan release notes: Version 11.8.0. https://documentation.cochrane.org/revman-kb/revman-release-notes-342261851.html

Nested Knowledge. (2026). Release notes: Versions 1.115.0 and 1.115.1. https://about.nested-knowledge.com/docs/releases/

Schunemann, H., Brozek, J., Guyatt, G., & Oxman, A. (Eds.). (2013). GRADE handbook for grading quality of evidence and strength of recommendations. The GRADE Working Group. https://gdt.gradepro.org/app/handbook/handbook.html

von Schreeb, L., Bruckner, T., Aspromonti, D. A., Caquelin, L., Cummins, J., DeVito, N. J., Axfors, C., Ioannidis, J. P. A., & Nilsonne, G. (2026). TrialScout links published results to trial registrations using a large language model. Journal of Clinical Epidemiology, Article 112484. https://doi.org/10.1016/j.jclinepi.2026.112484

Note. CRAN publication dates for metafor, meta, netmeta, robvis and metaumbrella were read from package metadata on 27 August 2026.