A manuscript is usually verified once, by reading it. Its numbers are checked against the analysis when they are written and never again, so a regenerated figure, a renumbered section or a rewritten paragraph can leave a published value pointing at nothing with no outward sign. Worse, a check whose anchor no longer matches the document finds nothing to object to and reports a pass while examining nothing at all. This record describes an architecture built against that failure and used through nine months of work on a gravitational-wave ringdown analysis: a registry in which each of 130 printed numbers is re-derived from the raw data on demand and its carrying sentence checked with it; a hash chain that refuses to run the verification suite until a change is declared; an append-only ledger in which retractions are appended rather than corrected in place; tamper tests that require each check to fail on its own specific defect; and the rule that anything derived about the document's structure is rebuilt by exactly one function that every check reads from. The record includes the full development history it describes: 313 files of ledger, plan, decisions, measured findings and item studies; a 51-link hash chain over both documents; and an inventory of 1,059 dated manuscript snapshots spanning 2026-07-25 to 2026-10-27, each with its SHA-256. Licences. The architecture source in verification_architecture/ is under the Apache License 2.0, as its own LICENSE file states. The method note, the paper and the development record are CC-BY-4.0. This record includes the architecture's source: 27 modules, 9,195 lines — the gates, the citation-fidelity audit, the mutation (tamper) suite, the document primitives, and the harness that runs them. It is readable but not runnable on its own: the posterior loaders and the pipeline they belong to are in the companion restricted deposit. The analysis engine itself and the manuscript are deposited separately under restricted access, all rights reserved.
No takes yet. Share an insight, caveat, or question.
Attia A. Gadallah (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: