Constraints on the recurrence times of subduction zone earthquakes are important for seismic hazard assessment and mitigation. Models of such megathrust earthquakes often assume that subduction zones are segmented and earthquakes occur quasi-periodically owing to constant tectonic loading. Here we analyse the occurrence of small earthquakes compared to larger ones-the b-values-on a 1,000-km-long section of the subducting Pacific Plate beneath central and northern Japan since 1998. We find that the b-values vary spatially and mirror the tectonic regime. For example, high b-values, indicative of low stress, occur in locations characterized by deep magma chambers and low b-values, or high stress, occur where the subducting and overriding plates are strongly coupled. There is no significant variation in the low b-values to suggest the plate interface is segmented in a way that might limit potential ruptures. Parts of the plate interface that ruptured during the 2011 Tohoku-oki earthquake were highly stressed in the years leading up to the earthquake. Although the stress was largely released during the 2011 rupture, we find that the stress levels quickly recovered to pre-quake levels within just a few years. We conclude that large earthquakes may not have a characteristic location, size or recurrence interval, and might therefore occur more randomly distributed in time.E lastic rebound theory, introduced first by Reid following the 1906 San Francisco earthquake 1 , is one of the foundations of earthquake science and explains how tectonic forces load faults. It states that tectonic stresses build up on a fault over decades, to be released within a major earthquake in seconds. However, it is still unknown if this release is complete and followed by a period of gradual reloading-and thus relative safety-or if sufficient energy remains in the system to allow similar size events more or less immediately. To explore this question, we use a fundamental observation in seismology, the exponential relationship between the frequency and magnitude of earthquakes, known as Gutenberg-Richter law 2 , log 10 (N ) = a − bM, where N is the number of events equal or above magnitude M, and a and b are constants. This relationship is commonly used to infer occurrence rates of infrequent large and hazardous events from the productivity level (a-value) and size distribution (b-value) of abundant smallto-moderate-magnitude seismicity. Although on a global average b ≈ 1, local b-values show substantial spatial variations-that is, in some volumes the proportion of larger magnitudes is higher (b < 1), in others the proportion of small magnitudes exceeds the average expectation (b > 1).Evidence from laboratory experiments 3,4 , numerical modelling 5 , and natural seismicity 6-8 indicates that b-values are negatively correlated with differential stress. Fault patches of such-determined significant stress accumulation have been observed to coincide with locations of subsequent large earthquakes 9,10 . Low differential stress conditions, for exampl...