Three Lenses on a Trustworthy Evidence Base: Completeness, Authenticity, and Robustness as a Reproducible Computational Integrity Audit

Abstract

Background: Trust in the clinical-trial evidence base is usually defended one study at a time, by hand. We ask whether it can instead be audited computationally, at registry and corpus scale, and whether such auditing can be honest about its own blind spots. Methods: We assemble three automated, reproducible lenses on a single programme and apply each to public data: (a) field-level structural missingness across the full ClinicalTrials.gov registry (AACT snapshot, 12 April 2026; 579,828 studies); (b) a Benford leading-digit screen of 1,175,056 numeric values across 403 Cochrane reviews (the Pairwise70 corpus); and (c) the Walsh statistical Fragility Index recomputed from a landmark trial public 2x2 table (DAPA-HF primary composite). Results: Completeness is poor and uneven: 62.9% of records lacked publication links, 48.2% lacked IPD sharing statements, 32.8% lacked a detailed description, and 10.2% lacked locations, with industry the most affected named sponsor class. Authenticity returns an honest null: the corpus first-digit mean absolute deviation (MAD) was 0.013 (95% CI 0.013-0.014), within Nigrini marginal-conformity band, with no corpus-level digit anomaly. Robustness is modest where it matters: a result significant at p = 1.6x10-5 rested on a Fragility Index of 62 events out of 4,744 randomised (Fragility Quotient 1.31%). Conclusion: Completeness, authenticity, and robustness form a layered, deployable integrity audit — but each lens is bounded, and naming those bounds is part of the method, not a footnote to it.