Answer coverage

Two results can carry the same label, run the same task list, and print the same headline percentage while resting on very different numbers of answers. Answer coverage separates them: it is the fraction of requests sent that came back with a scorable answer, reported beside any rate computed on those answers. A refusal, a timeout, a truncated reply, and a rate limit each remove a request from the pool before anything can be scored, and none of them leave a mark on the rate itself.

Report it as a fraction with both counts showing, per task family, next to the rate it qualifies. Three counts do the whole job: requests sent, answers returned, and answers correct.

Zero answers means an empty denominator. Print the request count and the reason beside it, and never publish the cell as a rate. An answer cut off by an output cap is its own kind of outcome. Give truncations a bucket of their own, apart from wrong answers and apart from refusals.

Coverage is the measurement that makes coverage conditioning visible, the way a rate comes to depend on which requests came back, and the two are published together. An interval is computed on the answered denominator, so an eval confidence interval says only as much as the coverage figure printed beside it. A three-model run that prints the answered fraction beside every score, one task family at a time shows what the practice looks like on a real record.