An LLM-driven evolutionary search, run three times against the map from Part 1. The same loop returned 0.96x when its fitness sat on a noise floor, a kernel worth 1.04x end-to-end when pointed at the wrong op, and, once measurement chose the target and kept the score honest, a fused attention kernel at 80% of the hardware ceiling that Apple's stack ran at 34%. Every campaign mined candidate by candidate: what each jump changed, what analysis provided versus what the search invented, and the verification gauntlet between a winning candidate and a shipped number.
Apple documents almost nothing about how its GPUs work: no instruction set, no latencies, no occupancy rules. Rather than reverse-engineer the chip, this measures it, then scores every operation in Apple's own inference stack against its own ceiling. The result is that MLX's matrix-multiply kernels already run at 85 to 102% of roofline while its attention kernels range from 33 to 92%, and the largest gap sits in prompt processing rather than in generation. Includes the eight ways a plausible measurement turned out to describe something the software never does.
How butterflies, bureaucracies, and coding agents fail in the same way
ETL fails in the same ways every time. Those ways are enumerable. Enumerated problems are solvable.
Anti-fragile, durable deep-research.
One-liner in, code-complete out.