We're building a shared corpus repository for our fuzz targets. Every time the fuzzer runs, it builds up optimized inputs that make future runs more efficient—it doesn't start from zero each time. Over time this becomes an automatically growing collection of edge cases that doubles as regression testing.