Parallelization

Run many LLM calls at once and combine them — sectioning versus voting, the latency math, aggregation as the hard part, why consensus is not verification, and when not to parallelize at all.