Bridging the QC–QA Cognitive Handoff: An Agentic Verification Harness for Error Detection, Specification Cross-Checking and Quality Documentation Assembly in Pharmaceutical Quality Assurance

Abstract Background. Quality control (QC) results reach regulatory dossiers through manual transcription and review. Such a benchmark does not yet exist. No collection of pharmaceutical quality documents carries discrepancies of known class and location, so automated verification cannot yet be evaluated against measured performance across realistic transfer errors and document types. Objectives. To specify (i) an architecture in which large language models only read and propose values, deterministic code performs every calculation and acceptance test, and a qualified reviewer makes every decision; (ii) a protocol for building a reference corpus and evaluating such systems; and (iii) sensitivity estimates from three dossiers. The plan covers design, evaluation, and measurement. Methods. Records from three ASEAN Common Technical Dossier Part II submissions (13, 14 and 10 records) were de-identified, and 112, 100 and 120 discrepancies across nine classes were introduced into dossiers A, B and C. A finding counted as a detection when it identified an injected discrepancy at the correct location with the correct class. The framework ran blind to the register. Sensitivity to injected discrepancies is reported per dossier with Wilson 95% confidence intervals and pooled by random-effects meta-analysis; clustering within records is not accounted for. Results. The harness detected 95 of 112 (84.8%; 95% CI 77.0–90.3), 80 of 100 (80.0%; 71.1–86.7) and 105 of 120 (87.5%; 80.4–92.3) injected discrepancies, a pooled sensitivity of 84.1% (72.4–91.5), with detection in every class in every dossier. False-positive rate, specificity, precision, run-to-run repeatability and performance relative to human reviewers or baseline systems were not measured, and the injection registers have not been independently validated. Conclusions. The architecture detected a similar share of injected discrepancies in three dossiers, one run each. Although single runs across three dossiers cannot show reduced reviewer workload or superiority over existing practice, the observed pattern supports further testing rather than any claim of practical benefit. The protocol defines the required next experiments.