Not tested is not negative: assay-level panel composition and the limits of driver-negative classification in non-small-cell lung cancer

Background. First-line therapy in advanced non-small-cell lung cancer (NSCLC) is decided by driver status and PD-L1 expression. A molecular report stating that no driver was found is routinely treated as evidence that no driver is present. These are different claims whenever the assay did not interrogate the gene, and clinical decision-support systems do not generally distinguish them. Methods. We implemented a deterministic NSCLC first-line decision engine in which every gene carries three states — positive, negative and not tested — and which declines to issue a recommendation when a required gene was not interrogated. Gene status was assigned from each sample's own sequencing panel rather than from a cohort-level assumption. The engine was run without modification across four public cohorts (n = 4,816) obtained through the cBioPortal API, and against outcome-bearing cohorts for threshold evaluation. All rule registries are content-hash stamped. Results. The engine produced zero errors and zero violations of its safety property (a driver-positive patient is never refused for incomplete testing) across all 4,816 cases, with byte-identical output on re-run. Panel composition, not institution, determined coverage: in one metastatic cohort, NTRK1 was uncovered in 1,364 of 2,621 patients (52%), and this was attributable entirely to a single 18-gene liquid-biopsy assay that does not carry NTRK1. Every tissue panel examined (IMPACT341/410/468/505, ACCESS129) covered all seven first-line targetable genes. In an independent cohort of 2,653 cases, all seven genes were covered in all 2,422 cases with a determinable assay. The PD-L1 ≥50% threshold on which the engine routes driver-negative patients separated progression-free survival in two independent immunotherapy-treated cohorts (n = 246: 2.2 vs 6.2 months, log-rank p = 0.00053, HR 0.583, 95% CI 0.428–0.794; n = 86: 3.5 vs 10.0 months, p = 0.0104, HR 0.462, 95% CI 0.252–0.848). Conclusions. Driver-negative status is a property of the assay as much as of the tumour. Because the genes omitted by focused assays are precisely those with the lowest prevalence and the newest matched therapies, systems that do not record panel composition will convert rare, actionable, untested genes into negatives at scale. We report no clinical concordance study; the protocol for one is pre-registered and content-hash locked.