TwinForge 90/5: Coverage Driven, Residual Calibrated Digital Twins with Certified Gap and Drift Bounds

Closed-loop autonomy requires sim-to-real validity under coupled conditions spanning environment, agents, sensors, and system timing.TwinForge-90/5 defines a contract for auditable digital twins using interaction-aware scenario coverage, paired gap bounds, and drift reporting.Scenario generation targets feasible interaction triads and mandatory critical motifs under an operational profile that supplies evaluation bins and weights.Simulator calibration uses a multi-fidelity twin augmented with learned Reality Residuals and an Active Realism Tuner that aligns taskrelevant feature statistics under validated SIL/HIL timing parity.Realism evaluation uses Counterfactual Twinning to construct matched real-synthetic pairs and reports match acceptance rates, distance distributions, and covariate balance diagnostics.The Gap & Drift Auditor converts perception, planning, and control KPIs into a unified per-bin risk, computes a Weighted Relative Gap with paired bootstrap confidence bounds, and produces Drift Cards reporting Earth Mover's Distance shift and a conformal risk inflation bound \delta_{0.95}atconfidence 1-\alphawith \alpha=0.05.Experiments on Waymo Open, nuPlan, and Oxford Radar RobotCar satisfy the contract with SIC_3=0.94,criticalmotif coverage =0.97, and WRG=0.036with95% confidence interval 0.0280.045.Factor ontology, binning, feasibletriad rules, critical-motif library, and KPI-to-risk mapping are fixed and included as contract objects to support repeatable gate outcomes.