I started the spreadsheet out of spite. In December I lost three online games in one evening to the same 1690 in three minute chess, and rather than accept the obvious explanation I decided to collect evidence. The obvious explanation turned out to be correct. The spreadsheet was still worth it.
Four hundred games between 4 January and 5 August. Here is what I counted, what I changed, and the several reasons you should not trust my numbers as much as I do.
I needed a definition I could apply the same way at 1am and at 8am, so: a move where the engine evaluation falls by 200 centipawns or more, my moves only, checked afterwards on a laptop that is old enough to vote in some countries. Not "a move I regret". Not "a move that lost the game". A number.
I also fixed a window. Moves 15 to 25, which is where my games are decided and where I now know 58 percent of my counted blunders live. Openings I know. Endgames I am slow but sound in. The middle is where I was throwing games away, so that is where I measured seconds per move.
One aside on labels, since it changed how I think about my own practice. FIDE's rapid and blitz regulations, the ones in force since March 2024, define blitz as more than three minutes but not more than ten. Three minutes with no increment sits below that line. So the games where I was worst were not even blitz. They were bullet, and I had been calling them practice for about four years.
| Time control | Games | My moves | Blunders | Per 100 moves | Seconds per move, moves 15 to 25 |
|---|---|---|---|---|---|
| 3+0 | 63 | 2,150 | 89 | 4.14 | 1.9 |
| 5+3 | 148 | 5,624 | 146 | 2.60 | 3.2 |
| 10+0 | 22 | 946 | 13 | 1.37 | 5.5 |
Nineteen hundred moves of three minute chess at under two seconds a move. I was not calculating. I was recognising, which is a different faculty, and it is the one that tells you a bishop on b5 looks right because a bishop on b5 has looked right before.
Two rules, both stupidly simple, both of which I broke constantly at the start.
The first: no 3+0. Not as practice, not as a warm up, not while waiting for food.
The second: on any move where material can change hands, either way, I name my opponent's best reply before I touch a piece. Subvocalised, lips moving, embarrassing on a bus. And a floor of four seconds per move between move 15 and move 25, enforced by looking at the clock afterwards and marking the game as broken if I dropped under it.
I logged the breaks, because a rule you do not measure is a mood. Forty one breaks in the first twenty games. Six in the last twenty.
| Time control | Games | My moves | Blunders | Per 100 moves | Seconds per move, moves 15 to 25 |
|---|---|---|---|---|---|
| 3+0 | 0 | 0 | 0 | n/a | n/a |
| 5+3 | 137 | 5,480 | 93 | 1.70 | 6.4 |
| 10+0 | 30 | 1,290 | 14 | 1.09 | 11.8 |
The 5+3 numbers are the only ones with enough games behind them to say anything. For a count of blunders over of my own moves:
Which gives 2.60 plus or minus 0.22 before, 1.70 plus or minus 0.18 after, and a difference of 0.90 plus or minus 0.28. A shade over three times its own error bar. Something moved.
The 10+0 rows are 13 blunders against 14. That is not data, it is an anecdote with a decimal point in it, and I am only printing it so nobody quotes it at me later.
There was a cost, and I did not expect it to be this size. I lost on time in 11 of 148 games before and 19 of 137 after, which is 7 percent going to 14. Two of those were positions I was winning. If you want the honest version of my summer, I stopped giving games away in the middlegame and started giving a smaller number of them away on the clock instead.
The design is bad and I knew it was bad while I was doing it.
I graded my own moves and I always knew which period a game came from. Nothing was blinded. If Priya showed me a client's numbers collected this way I would ask who checked them, and she has written a whole piece about the moment your own tidy dataset starts lying to you (@ledger/benfords-law-and-when-it-lies-to-you).
Cutting bullet meant I stopped playing at about half past ten instead of twenty to one. Sleep went from roughly five and a half hours to seven. I cannot separate the sleep from the rule and I am not going to pretend the rule gets full credit. If I had to guess I would split it evenly, and a guess is what that is.
I was also coaching more sessions, which means more hours spent looking at positions slowly with somebody asking me why, and that is training whether I file it as training or not.
My opponent pool drifted upward with my rating, which pushes the blunder count the other way, so that one is at least working against me.
And the rating, since somebody will ask: FIDE 2038 to 2051 across two rated events and 18 games. K is 20 for anyone under 2400, so thirteen points is a rounding error wearing a nice shirt. Tomas has a piece about measuring his own sky properly (@dust/my-sky-is-not-dark-and-i-have-the-numbers) and the discipline is the same one: your instrument is fine, your inference about what it means is where you will embarrass yourself.
Written for a 1400 who plays too fast, which is most of them.
The last item is the one everybody skips, and I think it is doing about half the work in the table above.