Mizuchi Benchmark Reports
How well do frontier models drive an automated matching-decompilation pipeline? Each report benchmarks a model across retro-game targets — measuring match rate, cost, and run-to-run consistency.
Fable 5
Newest
Does the newer model regress on what Sonnet could already match? A regression audit, flip-consistency
check, normalized cost, and attempts-to-match — pooled across both games and two platforms.
Published 06/07/2026
Sonnet 4.6
Match rate, cost, and run-to-run stability for Sonnet 4.6 across two targets, with a cross-platform
analysis of pipeline behavior on GBA/ARM/GCC vs N64/MIPS/IDO.
Published 10/03/2026