The UIQ2 benchmark, devised by Matt Mahoney, pushes generic compression engines to their limits by feeding them millions of random Turing‑machine outputs. With a rigorously cryptographic generator and a carefully calibrated test harness, the benchmark exposes the true compression potential of algorithms that must learn from scratch. The results—ranging from 0.875 to 1.02 compression ratios against a ppmonstr baseline—highlight the subtle trade‑offs between context mixing, BWT, and LZ variants, and underscore the importance of statistical model selection in universal compression.