Raycaster

Raycaster Eval publishes actual agent runs on real business tasks — reviewing warranty claims, editing a consulting slide deck, building merger-model sensitivities — and shows the full assignment, every agent step, the output file and the rubric result, not just a summary score. Each run lists the model and harness used, the exact run cost and input token count, and the specific rubric checks it passed, like stating a CPAM value to the cent. Some runs open into a live, read-only workspace so you can inspect the agent's actual output file yourself.