Data supporting "Optimizer-induced surrogate error across five model families: limits of risk-based release in building energy design"

This dataset contains the analysis-ready simulation data and numericalresults supporting the manuscript “Optimizer-induced surrogate erroracross five model families: limits of risk-based release in buildingenergy design.” The release includes: (1) the prospectively locked 1,200-candidatedesign manifest; (2) cleaned EnergyPlus result summaries and thecandidate-level analysis master table; (3) label-free risk features,candidate-level risk scores, and locked site–prototype and analysis-roleassignments; (4) model-specific optimizer-induced selection effectsacross five surrogate-model families; (5) cross-model meta-analysis,matched-control and sensitivity results; (6) probability-calibration,held-out risk-threshold, and 19-point risk–coverage results supportingGate A1 and Gate A2; and (7) source data supporting the main figuresand tables. EnergyPlus outputs are simulator-based reference evaluations and shouldnot be interpreted as measured-building observations. The public record contains analysis-ready datasets, result tables,protocol records, figure source data, a data dictionary, environmentand version information, and SHA-256 manifests. Model-training,candidate-generation, optimisation-orchestration, and EnergyPlusbatch-execution source code, trained models, weather files, and rawsimulation-output directories are not included.