I'm pretty sure this is why they broke up the data demographically. You could have 5,000 Sailors take the survey, but if (for the sake of argument) only 50 of them are Master Chiefs, the statistical validity of the survey in regards to Master Chiefs will be significantly poorer than the overall sample. That's one of the ways people like Gallup reduce bias; tailoring the sample to the known demographic data of the US population. Because without knowing the results in your survey in advance, you don't know how many gruntled people to include to offset those who are disgruntled. What they're trying to find out is who is gruntled and who is not! So they (theoretically) would need to attempt to proportionally match the number of men, women, black people, white people, brown people, gay people, straight people, old people, young people, and so forth to some degree FOR EACH COMMUNITY. And be proportional with rank and designator too. Or determine and justify which of those variables don't matter and why.I think that you're confusing the CLT with the inspection paradox; the former being the sample size needed to account for a reasonable standard error and the latter being the method of obtaining a sample that provides an unbiased estimate. You usually can't quantify bias. So while the Navy's survey might have a low standard error from the sample size, if it is centered on the wrong value for the estimates (ie biased because a disproportionate amount of disgruntled Sailors bothered to take the survey), it would still be wrong and no one will ever know how wrong it is.
Even the Gallup poll struggles with making sure that its estimates are unbiased. All sorts of weird existential factors affect bias in the estimate.
But without either significant cooperation from Big Navy (this is an unofficial survey) or access to PII that an unofficial survey wouldn't get, I fail to see how they could feasibly have targeted groups to ensure that X black male heterosexual aviators between the ages of 25 and 35 gave the survey Y responses. Feasibly, you'd have to have self-reporting and trust that the guy behind the keyboard was a 27-year-old straight black man. And OBTW, if someone from a small population answered "no comment" or equivalent for a question (or refused to take the survey), that'd be another confounding variable, as you'd need to adjust the sample size, find another person, or have padded the population to account for that.
Point is, you can play this game all day. I only had to take a 100-level stat course for my major, and even I know eliminating bias is a bitch. Here, they gave us the raw data so other people can crunch their numbers and break out the statistical significance of sub-populations.