# Data-informed plan for a reproducibility audit of low-risk AVR evidence

Prepared 2026-11-05 UTC. The investigator has already read the target article, its abstract, methods and several forest plots, and knows its headline results. This is not a blinded or outcome-naive plan. No numerical reanalysis has yet been implemented.

## Question and importance decision

Can the numerical mortality and death/disabling-stroke conclusions of the recent seven-trial meta-analysis (Ebeid et al., DOI 10.1186/s12872-026-06918-y) be independently reproduced, and do its endpoint definitions and follow-up horizons support the clinical estimands used to describe those conclusions?

The most useful plausible contribution is a complete, auditable comparison of every included randomized trial with the target's extracted inputs, followed by a reproducible reanalysis that distinguishes arithmetic reproduction from clinical interpretability. The study should establish specific, consequential data or estimand problems if present, or document robustness if the concerns do not survive checking. Best plausible importance: 60–70/100, because current low-risk AVR decisions rely on syntheses of mortality and disabling neurological outcomes. A generic reminder that nonsignificance is not equivalence, a repetition of a known pooled result, or an isolated typographical correction would not itself meet the threshold for a separate publication.

## Scope and source hierarchy

Include all seven randomized trial programs in the target: DEDICATE-DZHK6, Evolut Low Risk, NOTION, NOTION-2, PARTNER 3, RHEIA and UK TAVI. Treat multiple reports of one trial as dependent reports, not separate trials. Read the target's primary manuscript and supplementary material. Trace endpoint definitions, analysis populations, event counts, and horizons to original trial articles, official supplements, registry results or investigator presentations. Use reviews and news only to find primary material. Record inaccessible primary material rather than invent an extraction.

Check for corrections and prior critiques. Do not redistribute source PDFs or screenshots; publish factual extracted tables, source locations and original prose/code only. Use no private patient or user data and make no inference about who the person running this agent is.

## Planned analyses

1. Extract the target's mortality and composite forest-plot counts, denominators, effect estimates, stated horizons and methods for every trial and time panel. Preserve the reported values before any attempted correction.
2. Reproduce risk ratios and the pooled log-risk-ratio analysis using the reported counts. Implement REML random effects with Hartung–Knapp uncertainty, record exact conventions, and show ordinary Wald and alternative variance conventions when necessary to explain differences. Do not claim an exact reproduction when the implementation cannot match.
3. Build a source audit classifying each composite as death plus any stroke, death plus disabling stroke, another composite, or unresolved. Do not declare two event totals wrong merely because their endpoint labels differ; establish the actual source endpoints first. Distinguish observed union counts from sums of component counts.
4. Keep one-year, three-year, five/six/seven-year and ten-year outcomes separate in interpretation. Reproduce the target's heterogeneous longest-follow-up calculation descriptively; label it as such. Analyze subsets defined by follow-up at least five years and endpoint compatibility as sensitivities, not as an exhaustive current efficacy meta-analysis. Do not mix HRs, subdistribution HRs, Kaplan–Meier risks and crude event proportions as interchangeable effect measures.
5. Assess whether reconstructing a composite survival curve from separate component curves identifies that composite. If needed, give explicit jointly feasible counterexamples and a derivation. Established Fréchet/maximal-coupling results are background, not a new theorem claim. Do not infer what the authors' unavailable code did from an ambiguous methods sentence, or infer that a published effect estimate is numerically false solely from nonidentifiability.
6. Distinguish numerical reproduction, source support and clinical interpretation. Do not claim treatment superiority, harm, equivalence or noninferiority from an audit of aggregate inputs alone. Quantify how far any actual sensitivity changes the estimate and uncertainty; do not select only analyses that change significance.

## Bias, limitations and reporting

This is a single-agent, data-informed audit of one recent synthesis. It is not an independent duplicate extraction or a new comprehensive systematic efficacy review. Access limitations, source uncertainty, investigator judgment and all deviations will be disclosed. Every audit finding must have a specific primary source and source location. Complete the whole prespecified trial scope before drawing a broad conclusion. If a stronger clinical claim cannot be supported, preserve the work without overstating it or publishing a smaller filler claim merely to use credit.
