A Two-Layer Deterministic Retrospective Computational Analysis : Dose Intensity, Decision Latency, and the Separation of Information Failure from Execution Failure

Abstract
Background. Sub-protocol dose intensity in curative-intent immunochemotherapy is a replicated adverse prognostic factor, yet retrospective review of an individual course cannot ordinarily say whether an attenuated induction reflected missing knowledge, a prescribing decision, or an access constraint. We applied a two-layer deterministic counterfactual model to one completed course to make that attribution exact.
Methods. Case P, a 29-year-old man with germinal-centre B-cell (GCB) diffuse large B-cell lymphoma (DLBCL), stage IV-B, International Prognostic Index (IPI) 3, Ki67 80%, biallelic KMT2D loss-of-function, MYC/BCL2/BCL6 rearrangement-negative, received induction at a relative dose intensity (RDI) of 0.51, was declared refractory at Deauville 5, progressed on salvage, and received CD19-directed chimeric antigen receptor T-cell (CAR-T) therapy at fourth-line-equivalent disease burden. Four dated decision nodes were re-scored under four access arms (ACTUAL, FEASIBLE, IDEAL, ORACLE), with the shortfall decomposed exactly into decision, access, and information components; a second layer propagated the first-node correction through the subsequent course; overall survival (OS) was simulated at 40,000 Monte-Carlo draws per arm under a mixture-cure model.
Results. The information gap was 0.00 at all four nodes. The decision gap was +23.58 at first-line induction; the access gap was +1.97 at one node and zero elsewhere. The fitted cure model returned a cure probability of 0.23 at the delivered induction RDI of 0.51 and 0.76 at the protocol 0.95. Simulated median OS was 27.03 months for the actual course against 43.43 months for the counterfactual (+16.40). Three written recommendations present in the record — dose escalation, a conditional Deauville-5-to-CAR-T rule, and CAR-T endorsed concurrently by three centres — were unexecuted within their window or executed after 145 and 47 days. All eight engine invariants passed; gap signs were stable under 48 draws of ±20% weight perturbation.
Conclusions. Nothing required to choose better was unknown at any node. The losses in this course were losses of decision and execution, not of knowledge, and a deterministic, temporally gated decomposition converts that finding into an auditable attribution naming the failure mode and therefore the remedy. The record is censored at 10.3 months of follow-up.