This study tested how well flow-zone-indicator, FZI, rock typing predicts permeability in uncored wells, where its error originates, and at what averaging interval log-based permeability becomes useful for a reservoir model. Core data of 572 plugs from three wells of a mature sandstone reservoir were utilized. Each well was withheld in turn. Rock types predicted from gamma-ray, density, neutron and sonic logs were compared with direct regression of permeability on the same logs. The blind error was allocated to its sources by an order-dependent sequential procedure, and random, blocked and blind cross-validation were contrasted. In-sample coefficients of determination of 0.90 to 0.97 reflected internal consistency, and blind errors were close to one log10 unit. Rock-type misclassification was the largest error component in every well, and the FZI transform contributed least. With wells mixed but neighboring plugs excluded, accuracy stayed close to the blind value. Mixing wells therefore did not materially reduce the error, which is consistent with a mismatch between plug and log support and with limited predictive information in the logs. Direct regression outperformed the rock-type route, and averaging over 10 ft reduced its error by 15 to 56% in the two densely sampled wells, depending on the averaging rule and bin origin. The procedure allocated the blind error to its sources, separated well transfer from neighboring-plug leakage, and measured error against averaging interval. These values describe three wells in one reservoir segment, thus the transferable contribution is the diagnostic procedure rather than the numbers. The averaged quantity is the interval sample average of plug permeability, not a flow-equivalent cell permeability.