By measuring relative earthquake arrival times using waveform cross correlation and locating earthquakes using the double difference technique, we are able to reduce hypocentral errors by 1 to 2 orders of magnitude over routine locations for nearly 8000 events along a 35-km section of the Calaveras Fault. This represents ~92% of all seismicity since 1984 and includes the rupture zone of the M 6.2 1984 Morgan Hill, California, earthquake. The relocated seismicity forms highly organized structures that were previously obscured by location errors. There are abundant repeating earthquake sequences as well as linear clusters of earthquakes. Large voids in seismicity appear with dimensions of kilometers that have been aseismic over the 30-year time interval, suggesting that these portions of the fault are either locked or creeping. The area of greatest slip in the Morgan Hill main shock coincides with the most prominent of these voids, suggesting that this part of the fault may be locked between large earthquakes. We find that the Calaveras Fault at depth is extremely thin, with an average upper bound on fault zone width of 75 m. Given the location error, however, this width is not resolvably different from zero. The relocations reveal active secondary faults, which we use to solve for the stress field in the immediate vicinity of the Calaveras Fault. We find that the maximum compressive stress is at a high angle, only 13¿ from the fault normal, supporting previous interpretations that this fault is weak.