I use the SB-5, and Dottie's got it pretty well right.
Basically, most standardized tests use a basal-ceiling structure. You start at a certain point, somewhere that is supposed to be very easy for the client, so that they will get them all right. If they get some number in a row right, you assume that anything easier than that they'd also get right, so they get credit for all of those items below the basal. Then you work your way up. Eventually, as things get harder and harder, they're not going to get them all right any more. And eventually, they're going to get them all wrong. If they get some number in a row wrong, you assume that anything more difficult than that they'd also get wrong, so you discontinue the task, and they get no credit for any items above the ceiling. (You'll hear the word "ceiling" used in two related but different ways, it can be a little confusing.)
The rules for where you start, what constitutes a basal, what constitutes a ceiling, and other fine points, are all defined in the test manuals. You have to follow the rules -- that's part of what makes it a standardized test.
One of the many annoying things about the Wechsler tests (WPPSI, WISC, WAIS) is that they do not allow the tester to adjust the starting point, which means that testing gifted kids takes . for . ev . er. The Woodcock-Johnson tests allow the tester to estimate a starting point (they have charts to give you guidance on each test). I like that a lot, because I can be efficient with testing and not bore the kid. I tell the kids I'm starting pretty high, so that if we have to drop back, it's no great shame.
On the SB-5, there's a routing procedure. You do one verbal and one nonverbal subtest where you start based on age, backtracking as necessary if the client doesn't establish a basal. Then the raw scores on each of those direct you where to start the other four verbal and the other four nonverbal subtests. On each of those eight subtests, items are grouped into 6-point sets at six different levels. If you get routed to level four on the verbal subtests and the client does badly on one subtest at level four, you drop back and administer level three of that subtest, and you keep backtracking as needed (rarely need to go back more than one level) until the client establishes a basal. Then you go back to where you were and move forward until the client establishes a ceiling. The only thing that's weird is that the subtests are interleaved with each other -- you do four nonverbal subtests at level four, then those same four nonverbal subtests at level five, etc. Makes me feel like an octopus sometimes, juggling all of the materials. It probably does reduce the "which subtest came last when the kid was tired" problem, and I think it's supposed to be more true to the old style tests where there was really only one test that kept shifting item type.
