A Validation Blueprint for Trajectory-Based Neonatal Early Warning: Endpoints, Comparator Ladders, Leakage, Lead Time and Falsification

This paper turns the validation discipline of the author's neonatal monograph into a blueprint in which each design rule rests on a proposition with proof, and the size of each pitfall is shown by simulation of synthetic data. The central claim is that trajectory-based early warning must earn its place by demonstrating endpoint-specific improvement over simpler comparators. Under proper scoring rules, the gain from adding a variable layer equals the conditional information that layer contributes; but a fitted comparison also contains the repair of the baseline's own misfit, so a weak comparator manufactures apparent gain for a layer with no information. The nested comparator ladder—counts, burden, delay and recovery, connectivity and thinness, then control response and context—can be tested in fixed sequence at the nominal error rate without multiplicity adjustment, provided the order is prespecified and each test is valid. Sample size is limited by the number of infants, not recording time, because repeated windows from one infant inflate precision by a design effect. Window-level splitting, features computed from the horizon, overlapping windows, event-anchored sampling, a shared measurement channel and dropped censored windows each bias apparent performance in a known direction. Warnings unrelated to the endpoint achieve substantial event-level sensitivity and lead time purely by chance, so both must be reported against a baseline at the same warning rate. Route accuracy is capped by adjudicator accuracy unless model and adjudicators share errors. When expected loss does not increase with signal quality, abstention improves accepted-window performance without changing the model, so coverage must be reported. Before-after analyses show improvement from regression to the mean alone, threshold rules supply a discontinuity design, and effective treatment makes a correct model appear miscalibrated. A staged programme—retrospective, prospective silent, interventional—is proposed, with falsification criteria and a protocol template. No clinical data are used and no clinical claim is made.