Independent Close Source Eval

Open source benchmarks are easily exploitable, so we run an independent eval suite drawn from day-to-day work.