ReproPatch

ReproPatch turns a bug video into a frozen, hashed reproduction test, lets Meta Muse Wonder patch the code through sandboxed tools, then runs an independent evaluator the model never sees. In 28 real runs on 14 real VanJS bugs, 8 bugs were verified fixed. Four times the model said "Fixed" while its patch was incomplete and its own tests passed; all 4 times the hidden independent check failed and ReproPatch stored REJECTED, so there were 0 wrong "fixed" verdicts. The problem: AI coding agents often produce patches that are "almost right, but not quite" (the top frustration for 66% of developers in the 2025 Stack Overflow survey), and "the tests passed" is not the same as "fixed". When we asked Meta's open-weight Muse Spark Open to review the same patches blind, it accepted all 4 incomplete ones. Only executed checks the agent cannot read caught them. So in ReproPatch a verdict comes from executed hidden checks, never from a model's opinion. Everything is evidenced: frozen test, patch and bug-video SHA-256 hashes, a per-call spend log (USD 7.14 metered in total), and an offline verifier (python3 scripts/verify_evidence.py). Judge page with every run and recordings: https://igoreduardoai.github.io/repropatch/ Honest limits: 14 bugs from one library is a feasibility pilot, not a benchmark; the independent checks are adapted by us from the upstream fixes' tests; no customers or market validation yet.