I didn't vote. The reason is that neither really supports what I beleive, aka, both are somewhat true. AMD certainly began basing the PR on extrapolated TBird performance, although even then whether an actual TBird would perform at those levels is questionable due to various bottlenecks and whatnot. The early XP's were definitely faster than the early P4 iterations(aka Athlon XP 1700 > (early)P4 1700mhz), so at that time it was clear that AMD was sticking to their claims. However, as Intel improved the P4, the difference between the Athlon PR and the actual mhz of the P4 began to sync with each other. AMD might still be using the TBird theoretical performance, but due to this synching, it's hard to not think that the P4 actual mhz is actually the current standard.
Interestingly enough, I think the Barton is the clearest "evidence" that AMD may have switched to a direct P4 comparison for their PR. The reason is that the Barton 3000+ kinda looks correctly PR'd when compared solely against the P4 3000mhz, but once you add the TBred 2800+ into the mix, it appears that the Barton 3000+ couldn't possibly be based on the same tests that the TBred 2800+ were based on. Unless, of course, AMD uses some type of method not used by most Internet based websites.
So, AMD definitely based ealy Athlon XPs against TBird, but whether they still do I dunno.