Hey there
I am mixing the song and hunting the picks to keep all of them under Zero before mastering. The paradoxical phenomena is that after I fix all the picks for the part and play it few times from the nearest marker and they all under Zero, I then play the song from the beginning and the same picks are clipping again above zero. I stop, play from the nearest marker and have no clipping. Some time I have small clipping at 0.1 for one sample, some time the same part would clip several times.
 
Another difficult task for me is to tell what track is clipping. I mute them one at a time and it does not clip, but as a sum of all tracks it clipping. This should be more accurate. 
 
So if I play from measure 100 and 32 measures part has no clips but when I play from measure 1, the part at 100 measure is clipping, what should I believe? Thank you.   

Lower the volume. 

Well- that is obvious, but not a solution. This is pre-recorded data, there is no reason for it to deviate in levels. Can you explain it? Yeah, I can lower it and give up on some volume. But I don't want to lower the entire song down because I have picks in 2 locations.

You don't have any effects which may have a random element to them? 

I have Izotope Nektar on Vocals, it has breath suppression forward looking- I suspect it causing latency- can not use Kontakt when it is active... Can it do that? I would think- if it is the same audio data going though the same effect, it should react the same- it is all math and algorithms. Isn't it?

Well let's say you have an effect which has some kind of time based modulation (delay, chorus, phaser, that sort of thing). If the modulation speed or delay time is not synced with the DAW and is specified in absolute units like Hz or ms, then the resulting audio will be slightly different depending on where you start playback from. Whereas if you've specified this modulation time in musical units like 1/8, 1/6th notes etc then it is synced with the DAW and the effect will be the same every time. 
 
But if you don't have an effect like that on the track in question then I really don't know, I mean in my non-expert mind I can imagine the possibility that something with a look-ahead algorithm could give slightly different results depending on where you start playback from, maybe a tiny difference in the way the resulting math is calculated could result in a small difference in rounding which just pushes a sample over the threshold when the calculation starts from point A as opposed to point B. But then again, I have absolutely no idea what I'm talking about and should probably leave the answer to the egg heads among us....

vladasyn
Hey there
Another difficult task for me is to tell what track is clipping. I mute them one at a time and it does not clip, but as a sum of all tracks it clipping. This should be more accurate. 
 

(I hope I'm understanding the situation correctly.)
That is the normal behaviour. All the tracks must be so low that the sum of them (master bus) doesn't clip.
It's not a question of metering accuracy per se. Alone muting one track can drop the sum below zero. The sum is always louder than the loudest track.
Need to ask: By picks - do you mean peaks?
 
 "...I fix all the picks for the part and play it few times from the nearest marker and they all under Zero, I then play the song from the beginning and the same picks are clipping again above zero"
 
Here, are you talking about clipping in master bus or the track meter? I assume, the master bus.
When fixing an individual track, do you have the track soloed, and when listening from the beginning, you have all tracks on?
 
What kind of levels do you have on individual tracks and what does the master show?
Do you use any other method for fixing  the peaks than track volume automation (like automating FX outputs) ?
 
If the volume really varies in a seemingly random way, it's very odd, and there should be a reason.

Yeah, Vlada - the various feeds to the master bus are summed, so the actual solution IS to lower the volumes of what feeds the master bus.  There are all kinds of WAY more knowledgeable folks here on this subject - and I might not even have said the above correctly, but I run into this all the time, when I have a bunch of extra buses for things like splitting drum kit into groups of kit pieces going to a sub-bus (my toms have one, cymbals have one, etc.).
 
It is NOT like the old analog days, where you would normally run things pretty hot, to keep noise out, and all of that.  As long as you are getting a decent signal presence, there is all kinds of room in the digital realm that makes it all much easier to keep the noise out without running too close to unity that things end up clipping.
 
Bob Bone
 
 

Let's talk about the bigger issue here: before mastering your peaks shouldn't be hitting near zero at all! Preferably your PEAKS should be somewhere between -10 and -4. Definitely not higher. Don't waste time getting them near zero, that's not your job. 

If none of the tracks are clipping, I'd consider just pulling down the input gain on the target bus. I don't think it makes sense to have to pull down each of the tracks a bit when the bus input gain is right there as a one-fader solution.
 
I've also seen the behaviour where the meters show something different for the same audio when I start playback from different points. I'm not entirely sure why this is, but I expect it's to do with the way that loudness is integrated over time. I expect the only safe solution is to back off a dB or two to be sure.