Residual tension in tendons embedded in prestressed concrete girders cannot be measured with conventional elasto-magnetic sensors, because those sensors encircle the strand and must be installed before casting. This article evaluates a system that magnetizes an embedded tendon from outside the member and estimates the change in its total anchorage force from the local flux density. Three ferrite-core solenoids of 100, 200 and 300 mm outer diameter are placed on a common axis within a 300 mm envelope, and a board carrying 25 Hall and six temperature elements sits between the magnetizer and the concrete surface. Tests used one 2 m specimen containing twelve 12.7 mm strands in the target duct, at the nominal 80 mm embedment and 300 mm stirrup spacing; four loading–unloading rounds gave 40 acquired states. Estimating absolute tension failed to generalize across rounds for every regression model examined, the error exceeding the spread of the reference itself. A within-session difference representation gave a cross-validated coefficient of determination of 0.942 and a root-mean-square error of 169.0 kN, 13.5% of the 1252 kN full-scale reference. The binding constraint is therefore the signal representation rather than the regressor. Transfer across specimens, preservation of a reference between measurement epochs, and detection of small losses remain unverified.