We evaluate Sietch on ProgramBench — reconstruct source from behavior alone, then check that it passes the original program's behavioral tests.
Functional-equivalence success reconstructing source on the easy & medium ProgramBench split.
Each target is evaluated binary-in, source-out — no source access, no decompilation, no internet.
A pass means every behavioral test for that target passes against our reconstruction — not a diff, an equivalence.
ProgramBench: 200 programs, 248,000+ behavioral tests, graded easy / medium / hard.
Complete methodology and per-target results are available under diligence — get in touch.