The thing I got wrong at first: treating the score as a verdict. It isn't one. Two people reaching for the same obvious library on a scoped brief aren't cheating, they're converging — and on a tight problem, convergence is the expected outcome, not the suspicious one. Boilerplate makes it worse: scaffolding, config, and standard error handling are near-identical by design. So the signal has to be weighted toward the parts of a submission where a real decision got made, and the number has to sit next to the work rather than in front of it. A human still makes the call.