Unfix — we unfix your fix, to prove it
The second key for AI-written C/C++ and firmware. VERIFIED, REJECTED or NOT VERIFIABLE (when the evidence can't decide) — on your own tests and real hardware.
The second key for AI-written C/C++ and firmware. VERIFIED, REJECTED or NOT VERIFIABLE (when the evidence can't decide) — on your own tests and real hardware.

Your AI agent says it's fixed. Unfix doesn't write code — it reproduces the bug, applies the fix, re-runs the evidence, then pulls the fix back out and checks the bug returns. Then it gives one deterministic verdict: VERIFIED, REJECTED or NOT VERIFIABLE (when the evidence can't decide).
unfix · nucleo-demo
Watch it on real hardware
1 min 46 s · sound on · recorded under the working name cplus
01 · A REAL BUG
Unfix builds the firmware, flashes the NUCLEO board and runs the test. It fails 3 of 3: the counter should be 42, it's 41.
02 · A FAKE FIX
The "fix" turns unity_fail into unity_pass. CI would go green. Unfix says REJECTED and points to the line.
03 · AN HONEST FIX
Eight independent checks on the board — bug before, gone after 3 of 3, nothing else broken, back on revert. VERIFIED.
The problem
Memory errors, races and flaky failures are exactly where AI coding agents — and tired humans — take shortcuts. The agent reports success; the bug ships.
The crash disappears because ASan or TSan no longer runs — not because the memory error is gone.
An assertion is loosened or a failing case is skipped, so the suite turns green around the bug.
A sleep makes a race rarer on one machine. It still fails in the field, on the real board.
How it works
01
Build in a clean sandbox and confirm the bug fails on the base version — every run, not once.
02
Take the patch from the PR or the agent and re-run the reference tests with sanitizers on.
03
Remove the fix and make sure the failure comes back. If it doesn't, the test never proved anything.
04
Flag disabled sanitizers, weakened assertions and added delays in the diff.
VERIFIED — the bug is fixed, with evidence REJECTED — not fixed, or masked NOT VERIFIABLE — evidence can't decide; said so plainly
See it run
A firmware project on an STM32 NUCLEO-F401RE. Unedited screenshots of the terminal app.
01 · OPEN THE PROJECT
Project type, connected board, ST-Link probe and the sandbox are checked before anything runs. One key to start.
02 · REPRODUCE THE BUG
Firmware is built for ARM in a container and flashed to the board. The bug must fail every run — otherwise there is nothing to prove.
03 · CHECK THE FIX
Tests not weakened. Bug there before, gone after. Everything else still passes. And the bug comes back when the fix is pulled out.
04 · THE VERDICT
Apply it to your folder, make a branch, or export an HTML report. If the fix had cheated, you'd see REJECTED instead — like at the top of this page.
Where it runs
GitHub Action (demo)
In our demo repository, a GitHub check posts the verdict and the evidence right in the PR, so a reviewer can see whether the fix is real before merging.
The PR check runs in our demo repository today; partner integration is in preparation.
Real comments from our demo repository (posted under the working name cplus).
See the real pull requests: #1 (VERIFIED), #2 (REJECTED).
An MCP tool for Claude Code, Codex and other agents. The agent fixes; Unfix decides whether it's done.
A simple command-stream app with /bug, /fix and /check — for setup, investigation and demos.
Flash, run and check on your hardware rig — for teams that test on the board, not on a host.
Public benchmark
We took real bugs from well-known open-source C/C++ projects, let Codex CLI fix them, and had Unfix judge each fix against the maintainers' own reference tests — hidden from the agent. The rule for picking those tests is fixed before the agent runs (one documented exception, fmt-4960), and anti-cheat tests guard against gaming.
fmt spdlog zlib-ng cpp-peglib
Cases, patches and verdicts on GitHub4
VERIFIED
6
REJECTED
0
UNRESOLVED
Bug types covered: memory errors under ASan, a concurrent-flush race under TSan, edge cases and logic errors — including a pure-C case close to firmware code.
Firmware
Most embedded teams test on hardware. Unfix drives your existing flashing and test tooling on the real rig and separates outcomes that usually blur together.
Running today on STM32 NUCLEO boards via ST-Link.
FIXED The test failed before the change and passes now.
BROKEN The test passed before and fails with the change.
FLAKY ON THE BENCH The test is unstable without the change too; fix the rig before trusting any result.
FLAKY, WORSE It was flaky before, and the change makes it significantly worse.
For a single bug fix, /fix gives one verdict: VERIFIED, REJECTED or NOT VERIFIABLE.
Built for regulated teams
The verdict comes from builds, tests and sanitizers — no language model decides it. Deterministic rules, no LLM in the verdict. Flaky and race bugs are judged by repeated runs, not a single pass.
Each run leaves a report: what was reproduced, what was run, what changed. Useful for an IEC 62304-style change record.
Built for closed networks. No cloud, no LLM — Unfix runs entirely on your machines. After a one-time setup, builds and tests run with the network off.
Questions
Unfix is an independent verifier for bug fixes in C, C++ and firmware. It does not write code. It checks whether a fix — written by a developer or an AI coding agent — really removes the bug, and returns VERIFIED, REJECTED or NOT VERIFIABLE (when the evidence can't decide).
How does Unfix verify a fix?It reproduces the bug on the base version (it must fail every run), applies the fix and re-runs the tests with sanitizers, checks that other tests still pass, then removes the fix and confirms the bug comes back. It also scans the diff for masking: disabled sanitizers, weakened or deleted tests, added delays.
Why "unfix"?Because the strongest proof that a fix works is to take it out again: if the bug does not come back when the fix is removed, the test never proved anything.
Does Unfix use an LLM to decide the verdict?No. The verdict comes from builds, tests, sanitizers and hardware runs. Deterministic rules, no LLM in the verdict. Flaky and race bugs are judged by repeated runs, not a single pass.
Does it work with firmware and real boards?Yes. Unfix can build firmware, flash a real board (today STM32 NUCLEO via ST-Link) and run the tests on the hardware. /check gives per-test verdicts on your rig: FIXED (failed before the change, passes now), BROKEN (passed before, fails with the change), FLAKY ON THE BENCH (unstable without the change too; fix the rig before trusting any result) and FLAKY, WORSE (flaky before, significantly worse with the change). For a single bug fix, /fix gives one verdict: VERIFIED, REJECTED or NOT VERIFIABLE.
Where does Unfix run?As an MCP tool inside AI coding agents such as Claude Code and Codex, as a GitHub pull request check that posts the verdict as a comment (shown on our demo repository; not yet available for partner repositories), in the terminal, and on your hardware rig. Built for closed networks. No cloud, no LLM — Unfix runs entirely on your machines. After a one-time setup, builds and tests run with the network off.
What did the public benchmark show?We tested 10 AI-generated C/C++ patches against real bugs from open-source projects (fmt, spdlog, zlib-ng, cpp-peglib). Judged against the maintainers' own hidden reference tests, Unfix verified 4 and rejected 6.
Who is Unfix for?C/C++ and embedded teams of roughly 20–200 engineers, starting with medtech and other regulated products where "the agent said so" is not an acceptable answer.
Design partner program
We're working with a small number of C/C++ and embedded teams — 20 to 200 engineers — to run Unfix on their own code and hardware.
Texto de unfix.dev tal como lo publicó, leído de su página y mostrado aquí con el formato de El Autómata. El original, en unfix.dev · Volver a la noticia de El Autómata