How accurate is accuracy?

In terms of my ratings I don't find myself qualified in this discussion. But as a student of theoretical computer science with exposure to game tree optimization that falls in the domain of AI/ML and complexity theory, let me add a few thoughts.

Here Nicator65 is alluding to something important. Accuracy is measured against the benchmark of what the engine recommends. Therefore, accuracy measured by the engine is never absolutely accurate. It is only relatively accurate to shallower engines, and if it is as good as almost all other engines at the same depth would be considered to be pretty good. Since chess is a very hard game for a computer, it never computes the whole game. It looks at 3-4 moves ahead, and makes an empirical judgement. (In case you want to find out more: n X n chess is NP Hard, and in fact it is EXP Complete) But we humans often play from learned experience that we call instinct. We might see something that requires the engine to work for hours or days even. This argument might not age well, if large size quantum computers becomes available soon; it might also age well. We don't really know as yet, I think.

Therefore, a lot depends on how far (i.e. to how much depth) the chess engine was exploring. A move that makes material win might be giving you a better position where you can win the game. But it might not as well. If the engine is shallow, it might look like a blunder, but if it is exploring deep, it would find it a brilliant move. This often happens even at the highest level and deep engines.

An example that comes to my mind was at one of the games in the championship match Anand vs Gelfand. The engine kept reporting Anand's move as an inaccuracy until a few moves after when it changed the rating to innovation or whatever it is called; it was something the chess engine found out later that it was way better than the chess engine's own recommendation. It would almost always happen if you let a chess engine analyze a game between an aggressive player like Rudolf Speilman or Mikhail Tal versus a counter-attacking player like Paul Keres.  Here is such a game if you want to :https://www.youtube.com/watch?v=fLMaT7sb2vQ

Even at the lower level of rankings, there is a nonzero likelihood that, this is what happened with your friend. It is also possible that your friend has been learning very quickly after you played with him. there is also a significant likelihood that he was cheating. One way to have better confidence in your conclusion is to make him play more games. Because as humans, we are inconsistent and incomplete (shoutout to Kurt Gödel). And the more games we play, they cover more of our range of instincts, strengths and weaknesses. More games we play they look more random to us, and higher likelihood, we will find our weakness.

Im rated 800 and my BEST was 92.4 and im so proyd of It ( It was 38 movés)