When Models Outrun Their Evidence: Model Ambition and Data Support

How much complexity an evidence base justifies is a well worked question in environmental modelling, addressed through information criteria, Bayesian model selection, paired simple and complex models and structural error analysis. That literature is prescriptive and site-specific: each contribution says what a modeller should do for one system. None of it reports how often published practice honours the result, and the distinction matters, because a rule that is well founded and widely ignored calls for a different response than a rule that has yet to be written.

Geologic® did two things the existing literature does not. First, it consolidated the scattered criteria into one derivation, attributing each of twenty criteria to its sources and flagging the de-escalation triggers and the reversibility rule as the paper's own synthesis, reporting conflicts between sources rather than resolving them by assertion, and stating as judgement calls the four thresholds no source fixes. The criteria are arranged into three ambition tiers, from screening on sparse heads to distributed transient prediction with ensemble uncertainty quantification, and are used to separate the tier a study occupies from the tier its documented data support. Second, it set out to measure the gap: all 3,815 records of a defined corpus were screened against a protocol fixed before screening began, with two flagged amendments and a negative control of 494 records designated in advance that returned no includes. The prevalence estimate itself does not yet exist, and the work reports the design and the screening outcome rather than a rate the evidence does not support. Arid, data-scarce groundwater basins are the demonstration domain, because they isolate the problem: capable codes, thin observation bases, and depletion decisions that are irreversible in human timeframes.