How come a move which reduced my game score by 0.04 was called a mistake, yet others that reduce it by 0.8 are considered inaccuracies?

Doesn't make sense. Can you paste this engine generated analysis here? 

It has happened in pretty much every second game at least once. The best move will be, for example, +2, and my move is +1.9 (e.g.), yet it will call it a mistake. I'll find a good example.

what do those numbers mean?

An engine assigns a number of pawns ahead one player is - so if someone is +5, they're 5 pawns ahead - this changes between engines, based on how they rate things. Some engines may be more concerned with king safety, so if their king is less safe, they'll give a lower rating than an engine that is more concerned with a strong attack will (assuming you have a strong attack, that is).

Also, some engines do it so that +5 means white is 5 pawns ahead, whereas -5 means black is 5 pawns ahead.

ah ok, it seems chess.com takes all three into account

Oh, certainly. I'd bet that all strong engines take many, many more things than the few things I listed into account. The main thing that changes the analysis between engines is the way it factors in.

If the chess.com engine gives a .5 pawn advantage to a safe king, and a 2 pawn advantage to a strong attack, and you give it a position where white has a strong attack and a king that's uncastled, and therefore not as safe, it would show (assuming nothing else factors in) white is ahead 1.5 pawns. But if you gave it to an egine that had a 2 pawn advantage to a safe king, and a 3 pawn advantage to a strong attack, it would show white is ahead 1 pawn - for the exact same position. It's up to how the author weights them.

so at what point is the game considedred lost? about 3?

Its not the amount that it goes down, but rather the difference compared to the best move. I have gone from 4 to 7.3 and it is considered a blunder because I could have 11, or even mate. Its not just whether the move is good or bad, but how good or bad compared to your other possibilities

It depends what level you're playing at, and which engine is running. For example, I played Crafty-23.5 against itself, with white missing the king's bishop pawn. Evaluation was about -1 for Black, and it managed to win in a 5 minute per 40 moves game. But Crafty is a pretty good engine, and I imagine someone at a 1400-1500 level wouldn't be anywhere near able to show the weakness of that one pawn difference.

yeah, i read the each point is essentially worth 100 rating points, and 400 rating points is typically a given loss for the lower rated player assuming ratings are equal

This table shows the correlation between the difference in Elo points and the probability of winning a game (in bold). So for example if a player has won 60% of her games against an opponent, she would be approximately 72 Elo points stronger. Conversely, a player that is 150 points weaker than his opponent, has only a 30% chance of winning a game.

A 99% chance of a win is +677 rating points. Not sure how reliable these numbers are, but yeah.

Taken from here:

http://www.ascotti.org/programming/chess/elo.htm

makes sense to me, 400 rating points wins 92% of the time

I have had moves that are 0.04 less than the best move, and they were "mistakes".

In addition, is it possible to beat the hard computer (1628) in 2|1 while making mistakes on 51% (27) of moves? My bullet rating was 1300 at the time. I've lost games against the hard computer while making 3 inaccuracies throughout the entire game. I'm thinking the computer analysis is broken.

11. Ne5? has a score of 6.94 and was considered a "mistake".

11. e3 has a score of 7.07 and was considered the "best move".

This does not make sense. A difference of 0.13 between the best move and your move does not mean your move was a mistake. It probably means your move was the second best move.

Are you sure the advantage wasn't for your opponent? Your inaccuracy giving them a slight advantage, mistake giving them a bigger one.

I was white. After my move 11. Ne5? the score was 6.94. The analysis said the best was 11. e3 (7.07).