Structural ambiguity changes temporal transfer of stage–discharge relations in a tropical urban river

Continuous discharge records are usually inferred from stage, so their transferability depends on both calibrated parameters and the hydraulic formulation that maps water level to flow. We tested whether reported parameter vectors could be transferred independently of that formulation. Ten Generalized Likelihood Uncertainty Estimation (GLUE)-labelled and ten Differential Evolution Adaptive Metropolis (DREAM)-labelled vectors for the Meia Ponte River, Brazil, were propagated through two explicit reconstructions of a divided-channel Manning relation. The experiment used one supplied 25-point cross-section, 2,159 official daily stage–rating-derived discharge pairs at or below its lower endpoint during 2018–2023, and 199,207 transmitted 15-min timestamps. Under the reference bed-segment reconstruction, the DREAM-labelled median had lower root-mean-square error (RMSE) than the GLUE-labelled median (3.68 versus 7.88 m³ s⁻¹), but its deterministic ten-vector envelope covered only 2.59% of reference discharges compared with 94.07% for GLUE. A horizontal-layer reconstruction reversed the ranking and increased DREAM RMSE to 55.05 m³ s⁻¹. Uncalibrated stage offsets also changed error, bias, and coverage materially. Stage-standardised annual analyses retained opposite bias directions, and errors increased on days with larger intraday stage ranges. Thirty-two daily stages above the lower endpoint, including 17 above the maximum supplied elevation, were excluded from headline results and reported only as extrapolation stress tests. The transferable object was therefore the model–parameter pair rather than the parameter vector alone. Hydrometric studies that reuse hydraulic calibrations should archive the executable forward model and report datum alignment, structural sensitivity, directional bias, ensemble coverage, temporal resolution, and geometry exceedance together.