Model only clears harness friction while the fuzzer and the sanitizer keep the real score.

That is the right division of labor, even if it makes the demo less magical.

Someone still owns corpus care and the long run after the harness compiles.

Without those boring checks in the writeup, you get a green build and a story for the review board.

Glad they put real time into the checks that tell you whether the harness actually feeds the C record parser like a real caller would.

That is the part teams skip when a local model spits glue out in one shot and everyone wants to call it done.

Funny that the wall was always the boilerplate entry point.

Not the fuzzer, not ASan catching the stack overflow.

Just the unglamorous function nobody wanted to own on a Friday.

Quiet upvote for the offline draft path.

You can aim a local model at a C record parser without shipping the target off-machine, and that is the part that actually helps.

The rest is org incentives.

If green builds and AI-assisted tickets beat ASan findings from a long run, the model just speeds up the empty win.

Once the harness writes itself the bottleneck just moves.

Someone still has to fund the long run and own the ASan triage, and that is the seat the demo never staffs.

Quiet upvote for the offline draft path.

The model can write glue.

Someone still has to own the overnight ASan run.