Computer-assisted scientific evidence synthesis — systematic reviews, regulatory submissions, and policy-relevant literature summaries — is routinely produced from pipelines that ingest thousands of papers, classify them, extract structured findings, and aggregate results into a published claim. Reporting standards (PRISMA 2020, RECORD, FAIR, FAIR4RS) describe what authors must report, not what reviewers can independently verify. Under collegial peer review this is adequate. Under adversarial review — by regulators, opposing counsel, Cochrane assessors, or competing sponsors — it is not. An adversarial reviewer must determine, in bounded time without applicant cooperation, whether the corpus, extraction, analysis code, and result are exactly what the applicant claims they were at submission. We describe a system that produces a single archive bundle composing (i) a frozen content-addressed cohort manifest, (ii) a scientific pre-registration reference, (iii) an extracted-findings dataframe with per-finding evidence spans, (iv) extraction-model and prompt-hash commitments, (v) a reference-implementation commit identifier and container-image digest, and (vi) an executable replay script — with a file-level integrity manifest, two detached cryptographic signatures from distinct parties, and deposition to a DOI-minting repository. A downstream publication cites the DOI together with both public-key identifiers; any third party can retrieve the bundle, verify integrity and signatures in constant time, and replay the extraction in a clean container. Each primitive is well known; their integration into an adversarial-audit-grade bundle for scientific evidence synthesis is, to our knowledge, not previously described.
Kyle Hackbarth (Tue,) studied this question.