In build failures, one of these assertions sometimes fails, with values just outside the range deemed acceptable:

    // The average random meta value of the first document should be about 0.75.
    ASSERT_GTE(firstTotal / nTrials, 0.74);
    ASSERT_LTE(firstTotal / nTrials, 0.76);

    // The average random meta value of the second document should be about 0.5.
    ASSERT_GTE(secondTotal / nTrials, 0.49);
    ASSERT_LTE(secondTotal / nTrials, 0.51);

After a deep dive into probability theory, Max and I were able to figure out that (assuming a perfect PRNG) the first set of assertions should succeed at least 99.9999% of the time (six 9's). If we double the range, we should get to at least thirteen 9's.

A similar analysis can be applied to the second set of assertions, but it gets messy since it involves taking the convolution of two Beta distributions, which is not fun.