Validated but unconstrained: why total-clearance data cannot constrain minor elimination pathways, with a worked case in intravitreal drug delivery

Mechanistic models in drug development are validated against quantities that can be measured and then used to predict quantities that cannot. This work shows that the inference from one to the other is quantifiable, and often close to empty. When the validation target is a total clearance rate k = k_a + k_p and the prediction of interest is the share f = k_p/k carried by one parallel pathway, the elasticities with respect to any parameter entering only k_p satisfy E_observable/E_prediction = f/(1-f), exactly. Validation informativeness degrades in proportion to how minor the pathway is, and minor pathways are what such models exist to quantify. Determining f to +/-10% requires the total rate to be known to about f x 10%: 1.5% for the ocular posterior route, 0.1% for antibody uptake into a solid tumour, against measurement precision an order of magnitude coarser. The general form is a Fisher-information statement that also captures confounding. The case study is intravitreal pharmacokinetics. A published first-passage model (Lamirande et al., IOVS 2024;65:21) is independently reimplemented in open-source code, reproducing its retinal areas to within 0.2-1.0% and its mean first passage times to within 1-4%. The effective one-way permeability of a vitreous-retina boundary backed by an exchanging retinal compartment is shown to be the series combination of inner limiting membrane and retinal pigment epithelium permeabilities, not the smaller of the two. The two coefficients answer different questions - arrival at the retina versus net posterior elimination - and differ by 1.3-1.6x. Because a single Robin boundary sets both the exit probability and the residence time, the reported posterior fractions are ambiguous between them, and the vitreal half-life moves by only 6-11%, too little to distinguish the two. Which barrier is tighter is not determined by the data: bootstrap resampling of the original rabbit dataset gives P(p_RPE > p_ILM) = 0.59 for Fab and 0.43 for IgG. The resulting bias is nevertheless bounded, min(p_ILM, p_RPE) <= 2 p_series, with the data lying within 4% of that bound. A second structural check - the model's global feasible set - shows the cynomolgus monkey half-life is reachable only by forcing a posterior elimination fraction of 56-66%, against the 10-30% the same model reports. Modelling the shortfall as a uniform volumetric clearance yields a single flow of about 0.21 uL/min accounting for both a 50 kDa Fab and a 150 kDa IgG (difference -0.067 to +0.120 uL/min), required in the monkey and not in rabbit or human: a falsifiable signature of convective bulk flow, which the test supports. No new experiments are reported. All source data are published. All analysis code is archived at doi:10.5281/zenodo.21887689 and a single command reproduces every number and figure. Preprint. Not peer reviewed.