Static Coulomb stress transfer is increasingly automated over open catalogues and fault databases rather than built by hand for single events. We ask how much of such a result survives its own inputs. For Türkiye we resolve the stress imparted by 46 Mw ≥ 6 earthquakes between 1990 and 2026 onto 6,970 sub-segments of the mapped active-fault network, screening each across 25 parameter combinations. Three silent failure modes emerge. First, 44 records of the GEM Global Active Faults Database store a central rake 180° from the circular midpoint of their own bounds; all lie in two source compilations, one being the principal North Anatolian Fault strand south of Istanbul, where it reverses the stress change from −0.029 to +0.251 MPa. Second, anchoring ruptures on catalogue centroids rather than mapped traces displaces them by a median of 18.4 km; across 27 testable events this flips the sign on a median of 11.9 % of segments against 2.0 % for a rake ±15° and dip ±10° perturbation, and is the larger of the two in 20 of them. Third, sign stability rises from 74.3 % to 89.5 % between 1 and 50 kPa. A pseudo-prospective test against subsequent seismicity gives a segment-weighted AUC of 0.711 for Mw < 7 sources, but clustering leaves an effective sample of 42, one event supplies 38 % of it, an event-equal estimate gives 0.589 ± 0.112, and distance matching leaves 0.655 with an interval spanning 0.5. Screening improves the estimate by +0.016 (95 % paired bootstrap [+0.001, +0.051]; one-sided p = 0.020). The forecast evidence is suggestive rather than strong; the two input-integrity findings do not depend on it. Coulomb maps should be published with an explicit statement of which parts are interpretable.