It seems to me, superficially, that the strips (in the NLA window) are now set to “Replace” the influence of other strips (instead of, say, “Add”), and also that the strips themselves were not each defined from the same “neutral” starting position and frame.
Ultimately, each “Action” is: “a named set of IPO-curves, all of which must either be ‘Apples’ or ‘Oranges.’” That is to say, “when you ask the NLA-system to <<add | replace | subtract | multiply>> them together, the end-result must make sense.”
Be sure that you’ve defined each action, starting with “frame #1,” where the rig is in a Neutral pose. Be sure that each Bone that you alter is only being altered to move the Neutral rig in response to this particular Action. Don’t specify changes to any variable if the Action does not require it to be changed. (If you find that this isn’t the case, you can fix things in “Tweak mode.”)
This way, you can be sure that your “Action definition” is truly generic: that it really can be “applied anywhere at anytime,” which of course is exactly what the NLA-editor intends to do. It must: (a) influence only the variables that it needs to; (b) influence them relative to a “common, neutral, starting-position,” and © start at a common Y-position, “zero.”
When all of these “minor issues” have been covered, then it will “lead to a sensible result” for <<this strip>>, at <<any>> particular frame-number, to be (for all of its “included variables” at once …) <<added | replaced | subtracted | multiplied>> to ‘the status quo.’ "
Since (heh…) that is what The Computer is ultimately doing, that is what you must be sure to provide.