Show HN: Green tests are not proof – verifier for C/C++ and firmware fixes

Unfix is an independent verifier for C, C++ and firmware fixes, especially changes written by AI coding agents. Its central argument is that a passing test suite does not prove a bug is fixed: a patch can disable sanitizers, weaken or skip assertions, or add delays that hide a failure. Unfix does not write code or use an LLM to decide whether a fix is sound. Instead, it builds the project in a clean environment, confirms the original bug fails repeatedly, applies the proposed patch, reruns tests with sanitizers, checks that other tests still pass, scans for masking changes, and removes the patch to confirm the bug returns. It reports VERIFIED, REJECTED or NOT VERIFIABLE, with evidence for the result.

The system supports local workflows, AI agents through MCP, pull-request checks, and hardware testing; its current firmware demo builds and flashes an STM32 NUCLEO-F401RE board. A demonstration caught a fake fix that turned a failing test into a pass, while an honest fix passed eight checks and earned a VERIFIED result in 31 seconds. In a public benchmark, Unfix assessed ten AI-generated patches for real bugs in fmt, spdlog, zlib-ng and cpp-peglib against maintainers’ reference tests, verifying four and rejecting six. It is designed for C/C++ and embedded teams, including regulated organizations that need reproducible, auditable evidence and on-premises operation. Pull-request integration is currently demonstrated in a public repository, with partner integrations still in preparation.