An agent emulates an elderly clopidogrel trial: 53 of 62 SQL checks met, calibrated risk ratio point estimate rises to 1.274

Target-trial emulation can now run from a clinical question through executable cohort construction and statistical analysis, while keeping the protocol, code, and results on file for human review. Earlier work covered either one stage of the workflow or spread design, coding, and statistics across separate tools, with uneven support for early design and cross-stage interoperability. On synthetic PCORnet-style claims, the system completed an ASCET-benchmarked emulation and an ages 81–90 extension with 6,069 and 1,602 patients, retaining 17 main SQL files and all execution diagnostics.

Language-model components sit under deterministic workflow control, and a local secure-execution environment passes downstream agents only metadata, execution diagnostics, and restricted aggregate summaries, so patient-level records stay local. A study-design router first proposes one of three pathways and any reference trial for approval, then the protocol builder keeps reference definitions distinct from observational adaptations, with missing measurements requiring explicit proxy or omission decisions. Two reviewers scored benchmark extraction, protocol, proxies, covariates, SQL, and R against checklists: 53/62 and 50/62 assessable SQL checks met, 12/13 R checks met for both, and all 18 covariates rated relevant by both.

Calibration transfers standardized divergence between benchmark and randomized trial onto the extension population, changing the extension relative-risk point estimate. The uncalibrated extension risk ratio was 0.848 with an interval including 1; after calibration the point estimate rose to 1.274 with a sensitivity interval of 0.439–3.700, still spanning 1. Calibration inputs came from ASCET first-event counts of 52/499 and 54/502, giving an unadjusted count-based risk ratio of 0.969 (95% CI 0.676–1.389) under complete two-year ascertainment, with the primary calculation omitting benchmark–extension covariance.

The system is intended for research, not individual treatment recommendations. The demonstrated scope is one retrospective synthetic clinical case: an ASCET-benchmarked emulation, an ages 81–90 extension, and an exploratory calibration, plus a ten-question routing development set, with execution limited to SQLite, claims-style data, and a two-treatment analysis engine. The authors propose locking application versions, data, approved specifications, and analysis settings before testing multiple studies and repeated runs; evaluation of at least ten additional studies is underway and their results are not included here.

The authors state that no causal ground truth was established, and that cohort counts and treatment estimates demonstrate workflow outputs rather than patient-level cohort accuracy, independent numerical correctness, or causal-effect recovery. Nine of 18 benchmark covariates and ten of 18 extension covariates were zero for every patient, the extension chronic kidney disease flag was also constant zero, the ACE inhibitor/ARB set contains a therapeutic-class mapping error, and weighted imbalance remained for male sex and encounter count at 0.220 and 0.421 in the extension. Calibration assumes benchmark divergence transfers to the extension, the primary intervals omit cross-age-cohort dependence, and the human workflow is a descriptive comparator rather than a reference standard.