Analysis Evaluation lines

When analyzing a game, it shows a magnifying glass icon next to your move and the best move that lets you follow the line, which is nice. However looking through the line of the best move more often than not leads to a horrible evaluation and includes inaccuracies and blunders the closer you get to the end of the line. Shouldn't they be the best moves?

I understand that the initial depth towards the end gets lower compared to the depth that analyzes each move, but then maybe the two top moves shown that let you explore the lines need to be calculated to as deep as the line and the "regular" depth added to that depth, so that reaching the line still displays the best moves.

For example in this game: https://www.chess.com/de/live/game/3520353612

Actually something even weirder just happened. On move 15. I played Qh4+ and it showed it as my only mistake

and said that Qc7 was best. I went through that line and it ended horrible, so I started writing this thread, wanted to go back to write the evaluation number that the mistake and best line end up as, but now, well, take a look

I mean, I wouldn't mind if my opponent took his own queen with his bishop, but I'm not sure that's how it works...

After refreshing I was able to run it again. The mistake line leads to +0,86, the best line leads to -4,99 but that's because the last move is a giant blunder that loses the rook for a pawn and the engine shouldn't include moves like that in an evaluation line

I'll have our devs take a peak and see what they think. 

As an interested layman I've glanced at all this and maybe this is Chess fighting to retain itself against the Comps. 

Capstoned by a recent book I bought, "I'm not crazy", "Today's Learners" are way ahead of "my old days" per rating, because of these instant (even 'half an hour is instant') reviews. 

But there used to be sophisticated tricks snuck into the various older engines that "lost cred against Stockfish and Houdini so fell out of favor", but some of them used "anti Master Algorithms" where if a line was chewed up by a 2-fold repetition "because it could", then "ran out of depth", they deleted the 2-fold and parked it and kept searching. 

Sometimes it was semi automatic, other times you could personally tweak engines that might not take the Comp Championships but you could purposely OverValue the Queen so it would Quit Sacking it for Junk Rooks, and the famous one was "Vicious Contempt for Draws" because "People Don't Give You Draws and Did You Notice Your Clock is Ticking?"

You could increase the value of knights just a hair and extend raw search by a dangerous extra ply and stuff. 

And Engines had "Personalities". We're all "getting Analysis" by one engine, and long long term that can't be good. Other engines might try something else and I think I recall from my glances at the field, first Rebel then Deep Junior had ways to utterly run Tal Settings at an annihilistic cost to all kinds of other evals but it enabled attacks you never heard of, instead of Stockfish's "Engine Champ Tuned" stuff. 

I only had one other training session against a guy here and two in a park where the opponents went for Nuclear.