Promoting Sponge (Water-Sensitive) Buildings: A Tripartite Evolutionary Game Analysis with Explainable Global Sensitivity and Predictability Diagnostics

Sponge and water-sensitive buildings carry a positive externality: incremental costs fall on developers and homebuyers, whereas flood-mitigation and ecological benefits accrue to society, so neither side adopts them spontaneously. Whether penalties or subsidies are the more effective remedy remains disputed. This study builds a tripartite evolutionary game of government, developers, and homebuyers at the building scale, in which enforcement intensity reg modulates both the expected penalty reg·F and the administrative cost reg·Cg, and estimates the sensitivity of adoption to the four policy levers within an explicit six-cell design that scans three government accounting bases and four structural regimes through a common baseline rather than crossing them. The central finding is that the comparison is not identified without stating the regulator’s accounting basis: under identical behavioural mechanisms, recording fines as government revenue makes the fine dominant (ST = 0.686), whereas recording fines and subsidies as internal social transfers makes the supply-side subsidy dominant (ST = 0.712 against 0.192 for the fine). Where homebuyers’ willingness to pay falls below the premium they bear, adoption collapses to a success rate of 0.031 and the ordering of three of the four levers ceases to be identifiable. The relevant deterrence variable is the expected fine reg·F, whose transition band remains at 0.014–0.034 of the baseline profit across profitability margins, while the margin raises the zero-deterrence success rate from 0.175 to 0.385. Replacing the unbounded conditional distribution of the surrogate with a Beta distribution matched to the bounded, bimodal response yields prediction intervals contained within the support of the outcome and a continuous probabilistic skill score of +0.014, against −0.051 for the Gaussian specification. Throughout, the response is the adoption level attained at a finite evaluation horizon rather than an asymptotic equilibrium, and its distribution is shown to depend on that horizon.