Local Factory
A software workshop running on my own GPU: a specification goes in, a diff comes out, the tests decide — with no call to any online service.
A harness that puts open-weight models running on my own machine to work. A task walks down a ladder of models: the smallest one tries first, and only failures climb a rung. Every attempt is checked against a specification and a test suite before an automated reviewer accepts or rejects it.
What interests me here is not the demo, it is the measurement. The project keeps a register of optimisations where my own ideas are marked falsified when the numbers say so — two of them are, and they stay on the record. Another time the harness reported five completed steps out of five in front of an empty diff: three separate defects, all leaning the same way, the agent’s way.
The guardrails are executable rather than written on a wall: the agent cannot commit to the main branch, and cannot push while the tests are red. That rule is a scar — an agent once wiped my working changes.
My conclusion after all of it, and it does not flatter me: the limit is not the harness, it is the models my machine can run.
Private project, no public link