How the analysis works
Every game played on Kliwe is read move by move by the engine. Here is exactly what happens, and what it can and cannot tell you.
The short version
The engine does all of it. There is no language model anywhere in this: no key, no call out to anybody else, no bill per game. Every sentence on an analysis page is assembled from something that was measured on the board, which is why the same game always reads the same way.
Step by step
1. The game is stored with a position for every move. Not just the notation: the full board after each half-move, so nothing has to be recomputed later and nothing can drift.
2. Every move is compared with the best move available. For each position the engine searches all the legal moves and scores them, then looks at what was actually played. The difference between the two is what the move cost, in pawns. That number is the whole basis of the analysis: a blunder is not a matter of opinion, it is a move that gave away three pawns when something else did not.
3. Book moves are left alone. While the game is still inside an opening line the book knows, a move is a book move and nothing else. Theory is not the players' work and a shallow search is no judge of it.
4. The moments that matter are looked at again, deeper. The whole game is swept at 2 plies, which is quick. The handful of moves the write-up will actually quote are then searched again at 3 plies, so nothing is called a blunder on the strength of a shallow look.
5. The write-up is assembled. The opening and how long it was followed, what each side gave away per move, where the game turned and why the better move was better, what to take away. Every one of those comes from the numbers above. Where a game has nothing interesting in it, the analysis says so instead of padding.
6. It becomes a thread. The analysis is posted to the community with a question at the end, because the interesting part is the disagreement.
What this is good at
- Finding the moment a game turned. It is usually earlier than you remember, and the number says so.
- Naming what went wrong in words you can use. "The bishop on c4 was undefended" is something you can check for in your next game. "White was better" is not.
- Counting honestly. You get what you gave away per move and how many inaccuracies, mistakes and blunders each side made. There is no accuracy percentage, because there is no honest way to turn those numbers into one.
What it is not good at
- Deep tactics. The search is a few moves deep, not twenty. A long forcing line can go past it. If the analysis calls a move a blunder and you can show the follow-up holds, you are probably right — and the thread under it is the place to say so.
- Plans. It can tell you a knight was hanging. It cannot tell you your pawn structure called for an attack on the other wing.
- Precise evaluations. The number is a measuring tape, not a verdict. Read "+1.4" as "better", not as "1.4 pawns better".
Playing with the tutor
The same engine will read a game while you are still playing it. Switch the tutor on next to the board and on every one of your turns it says what it would play, why, and what the move you chose was worth. It is the analysis above, arriving a move at a time instead of at the end.
Your games
Games you finish on the board are saved with their own page and read by the engine. Games you abandon are not stored. Saved games are public — that is the point of the archive — so play accordingly.