Rank-Deadline Guarding for Risk-Limited Kinase Assay Ordering

Historical activity profiles can diversify kinase assay orders, but their cross-panel utility is uncertain and they can postpone a high-probability assay arbitrarily far. We introduce rank-deadline guarding, an objective-agnostic wrapper that constrains a proposed order so every kinase stays within $K$ positions of a trusted marginal rank. Instantiated with greedy weighted profile coverage, it gives a deterministic first-hit delay bound, requires no training, and runs on a CPU. With $K=9$, it retained 80.6% of an unconstrained coverage gain under PKIS2 chemotype exclusion (+0.0433 discovery-curve area, 95% CI [0.0321, 0.0545]), while all eight development-condition estimates versus marginal ranking were nonnegative. We then froze the method and tested a separate PKIS1 panel after removing eight structures shared with development. Mean gains did not transfer: the external 80% threshold effect was -0.0206 (95% CI [-0.0306, -0.0117]). Across 4,428 held-out compound-condition evaluations, the implementation verified the theorem's consequence: zero delays of ten or more assays, versus 163 for unconstrained coverage. On external PKIS1 it was also consistently less harmful than unconstrained coverage. Thus the method solves a tail-risk problem, not universal ranking superiority: the untouched failure delineates when retrospective diversity is not evidence of translatable counter-screen benefit.