OK, the reason I could tell the difference so easily, as is the case in so many experiments, I had a flawed assumption, and mis-prepaired the samples, so what I thought I was testing I was not, but instead testing something else entirely.
I've had the discussion about sampling rates numerous times, and had even done the math to show what I believed to be correct. But as I said above, what I believed was not what the Nyquist Sampling Theorem actually says.
I always heard talk of the Nyquist limit, that being half of the sampling rate. That's not the whole picture. In my own calculations I saw trouble if a signal was right at the half-way point (like encoding a 24kHz sine wave at 48kHz), or if there was any harmonic content above the limit.
Turns out Nyquist had already accounted for these cases, and specifically excludes them from the theory. From Wikipedia, "the theorem shows that a bandlimited analog signal can be perfectly reconstructed from an infinite sequence of samples if the sampling rate exceeds 2B samples per second, where B is the highest frequency of the original signal." There's a lot of information in that little blurb. One, the sampling rate must exceed twice the highest frequency. So the 24kHz wave can't be perfectly reconstructed if you sample it 48,000 times a second, it must exceed 48k. The other, the signal must be bandlimited. That means the signal must contain no content at or above half the sampling rate.
So my test of a saw wave at 10k, was not valid. I picked that waveform because of it's harmonic content so there'd be something to hear in the upper octaves. It contains both even and odd harmonics. But that means the 10k fundamental had considerable content at 20k, 30k, 40k, and beyond. The first odd harmonic (30k) had already exceeded the limits of what the Sampling Theorem stipulated.
What happens when you have content beyond the Nyquist limit when sampling, is you get strange aliasing artifacts. It's a 1D interference pattern, you get peaks where they shouldn't be. Choose a different sampling rate and you get a completely different pattern of bad data, but it all sounds similar enough to saw-tooth harmonics that if you were to play only one sample it'd seem reasonable enough that what you were hearing was right. But when you sample exactly the same data at a different rate and get you get different harmonic content. It's obvious something is wrong.
So yeah, my ABX trial was fine, the data I was testing was produced by a incorrect premise. I'm trying to come up with a valid trial, but I need a lot of harmonic content to be audible, but can't have any of it reach or exceed 24kHz.
