Fernwood Espresso Works is invented, and so are its customers, its firmware and its knowledge base. The agent, the tools, the workflow, the scorers and the trace are the real framework, running in a Cloudflare worker. Pick a ticket and watch the run.
One of the three recorded runs below passes the judged scorers and cites a document it never opened. Only the trajectory shows it, which is the whole argument for the scorer that costs nothing.
| Scorer | Score | Why |
|---|
Three runs of this desk, written by hand as fixtures so the eval always grades the same three trajectories. Pressing the button runs the four scorers over all three, for real: the two judged scorers call the model, the two code scorers read the trajectory. The judged numbers move a little between runs, because a judge is a judge. The code scorers return the same answer every time, which is most of the argument.