Calibration Is Not Enough: TAP-Bathy v1.0, a Three-Axis Protocol for Evaluating Spatial Uncertainty in Bathymetry

Calibration is the standard yardstick for spatial uncertainty quantification (UQ), but it is insufficient. On the Amazon Fan, ordinary-kriging variance achieves excellent calibration (ECE = 4.1%) yet carries negligible spatial information(rpooled = +0.16). On the Blake Escarpment, kriging variance yields below-random block prioritization under the locked evaluation protocol. Motivated by this disconnect, we propose a three-axis protocol for evaluating spatial UQ: calibration, an exact within/cross-patch decomposition of the error–uncertainty correlation, and operational prioritization at a fixed resurvey budget. Applying the protocol to deep-sea bathymetric reconstruction across three terrains reveals three failure modes that any single axis misses – calibrated-but-uninformative uncertainty, statistically significant but operationally negligible withinpatch signals, and cross-region ranking reversal via Simpson’s paradox. Learned σ carries reliable cross-patch information(rcross = 0.08–0.50, locked evaluation protocol; confidence intervals at patch level) and concentrates 1.5–3× the random errorat 10% block-level resurvey budgets and remains above-random at 25%. A deep ensemble produces a near-identical primary endpoint, with the broad failure-mode pattern robust to the epistemic-uncertainty estimator, though secondary operational capture is unstable on the small Region 009 evaluation set. A pre-registered prospective test on an unseen fan/plain region yielded rcross = 0.358 inside the predicted interval, consistent with retained cross-patch informativeness on new terrain. An eight-region prospective test cleanly falsified the strong (sign-reversal) form of a terrain-coupling prediction. The contribution is a reproducible, scale-aware evaluation protocol: we show how to assess spatial UQ, not that one estimator universally wins.