I don't know whether this is a problem of the manual or the current implementation, and the way SLV works has been discussed recently, but anyway:
SLV seems to work better than in previous versions. For instance it no longer throws an error if the best possible function result exceeds 1E16. However, either the implementation or the manual is wrong: the latter states that the user provides two guesses and "for the rest, the user interface is as in the HP15C". It's not. And since the algorithms in the 15C and 34s are completely different, also the results differ. So this reference is misleading.
Example (tested on 3.1 3225):
LBL 99
X^2
2

RTN
FIX 4
1 [ENTER] 2 [SLV] 99 => X=1,4142 Y = 1,4166 Z = 1,4142 (= X)
Do the same on a 15C (or 34C) and X and Y will contain the two best possible solutions 1,414213562 resp. ...63 while Z holds f(X) = 1E9.
Or try the previous example with initial guesses 2 and 3. This will throw a +infinity error on the 34s, while the classis HPs return the correct result.
So the 34s obviously does not return the same results, and unlike the 15C/34C it also does not return f(X) in Z (which in this case would be approx. 9E11). The accuracy of the result depends from the display setting, i.e. FIX 4 and ALL will return different results. These are only as exact as the display setting, while the 34C/15C always return the best possible root(s). This behaviour should get documented and the 15C reference should be omitted.
Please don't get me wrong: SLV works well, but it is completely different from the 15C/34C. And yes, I would like to get results that are independent from the display setting. But the essential point is that the manual and the way the 34s works simply do not agree.
Dieter