2,627 rated blitz games · the whole archive, 100 days · 2026-05-23 to 2026-08-31 · current rating 2709
Nothing cleared the bar. No opening line, tilt pattern, session effect or colour asymmetry was both statistically significant and large enough to be worth acting on. On this sample that is a real result, not a gap in the analysis — it means the losses are distributed rather than concentrated in a fixable situation.
These are real: each is significant with the game as the unit of analysis. None carries an Elo figure, because none has a counterfactual measured inside your own games — you cannot elect to have blundered less, so there is no second group to compare against. A number here would be invented, and an invented number that looks measured is worse than none.
Your blunder rate is 2.7 per 100 moves normally and 11.1 on the very next move after a blunder — 4.1× the risk, on 2,769 such moves across 2,614 games (95% CI +7.2 to +9.7 points, p < 0.001). It decays but is still measurable 6 moves later, which was as far as this looked. This is the one thing on the page you can act on inside a single game, and it costs no study time: after a move you dislike, stop and look again before playing the next one.
With more than 80 seconds left you blunder 1.5 times per 100 moves. Under ten seconds that becomes 9.6, a 6.6× increase. Some of this is that hard positions eat clock as well as causing mistakes, so it is not purely the hurry — but the size of the gap is not explained by that alone.
Across 22,091 tactical moments you find 69% of forced mate chances (95% CI 64–74) against 97% of hanging piece chances (96–97). The intervals do not overlap. This is the same pattern as the winning positions you dropped, where the moves you missed were the quiet ones: what is obviously material you take, what has to be seen you sometimes do not.
When the tactic is forced — a recapture, a move with one sensible answer — you find 96%. When it has to be seen, you find 74%. Yet you spend 3.36s on the genuine ones against 3.07s on ordinary moves: 0.29 seconds more, on the positions that decide games. Nothing in your clock use distinguishes the critical moments from the rest, which is exactly the habit the blunder finding above also points at.
With 15 or more pieces on the board you blunder 2.0 times per 100 moves. At 9-10 pieces that rises to 5.6 — 2.8× — across 107,019 analysed moves. Piece count is a cleaner axis than the phase label, which is only a boundary rule; this is the same weakness stated in terms of the thing that actually causes it.
You blunder 5.4 times per 100 endgame moves against 2.3 in the middlegame (+3.1, p < 0.00001, across 1,826 games that reach both). That is the reverse of the usual pattern and it points study at technique rather than opening preparation.
Monthly mean rating runs from 2517 (May 2026) to 2704 (Aug 2026) across 4 months with at least five games.
Splitting the window in half, the second half is worse by 3.1 points of score (1313 games against 1314, p = 0.068, not significant).
| Line | Games | Score % | Opp Elo | Perf Elo | Edge | p | q |
|---|---|---|---|---|---|---|---|
| Caro-Kann Defense | 177 | 68.6 | 2557 | 2693 | +5.5 | 0.080 | 0.231 |
| Ruy Lopez | 248 | 67.5 | 2557 | 2684 | +3.8 | 0.149 | 0.231 |
| Sicilian Defense | 484 | 65.7 | 2565 | 2678 | +2.2 | 0.258 | 0.309 |
| French Defense | 147 | 57.8 | 2547 | 2602 | -6.6 | 0.076 | 0.231 |
| Pirc Defense | 45 | 57.8 | 2539 | 2594 | -6.7 | 0.341 | 0.341 |
| Scandinavian Defense | 50 | 56.0 | 2545 | 2587 | -9.4 | 0.154 | 0.231 |
These are openings reached while playing white, named by the opening rather than by whose choice it was: as White you appear here under your opponent's defence. Ordered by score %, the number you feel at the board — but read it beside Opponent Elo, because a line can score well simply because weaker players reach it. Where score %, performance and edge disagree, edge is the one to trust: it is the only like-for-like comparison. q is the Benjamini-Hochberg false-discovery rate across the 6 lines tested here; testing that many guarantees some p < 0.05 by chance, and only lines with q < 0.05 survive that correction.
After false-discovery correction, none of these 6 lines is distinguishable from par. Any ordering you see above is within what chance produces at this sample size — which is itself the finding: the repertoire is not where the points are going.
| Line | Games | Score % | Opp Elo | Perf Elo | Edge | p | q |
|---|---|---|---|---|---|---|---|
| English Opening | 83 | 71.1 | 2573 | 2729 | +8.1 | 0.081 | 0.484 |
| King's Indian Attack, with Bf5 | 54 | 68.5 | 2573 | 2708 | +7.4 | 0.220 | 0.562 |
| Sicilian Defense | 622 | 64.5 | 2557 | 2660 | +1.0 | 0.562 | 0.562 |
| Queen's Indian Defense | 142 | 63.0 | 2561 | 2654 | -2.4 | 0.528 | 0.562 |
| Nimzo-Indian Defense | 114 | 62.3 | 2553 | 2641 | -2.6 | 0.543 | 0.562 |
| Indian Defense | 93 | 58.6 | 2571 | 2631 | -3.5 | 0.466 | 0.562 |
These are openings reached while playing black, named by the opening rather than by whose choice it was: as White you appear here under your opponent's defence. Ordered by score %, the number you feel at the board — but read it beside Opponent Elo, because a line can score well simply because weaker players reach it. Where score %, performance and edge disagree, edge is the one to trust: it is the only like-for-like comparison. q is the Benjamini-Hochberg false-discovery rate across the 6 lines tested here; testing that many guarantees some p < 0.05 by chance, and only lines with q < 0.05 survive that correction.
After false-discovery correction, none of these 6 lines is distinguishable from par. Any ordering you see above is within what chance produces at this sample size — which is itself the finding: the repertoire is not where the points are going.
Game-to-game variation in edge has a standard deviation of 44 points, which is large: a single game is won or lost, and one result tells you almost nothing. That sets a floor on what any number of games can resolve. The column below is the smallest edge that would have reached p < 0.05 in each line, given how many games you actually played in it.
| Line | Games | Observed edge | Detectable | Visible? | Games needed |
|---|---|---|---|---|---|
| Sicilian Defense (black) | 622 | +1.0 | ±3.4 | no | 7148 |
| Sicilian Defense (white) | 484 | +2.2 | ±3.9 | no | 1491 |
| Ruy Lopez (white) | 248 | +3.8 | ±5.4 | no | 493 |
| Caro-Kann Defense (white) | 177 | +5.5 | ±6.4 | no | 240 |
| French Defense (white) | 147 | -6.6 | ±7.0 | no | 168 |
| Queen's Indian Defense (black) | 142 | -2.4 | ±7.2 | no | 1262 |
| Nimzo-Indian Defense (black) | 114 | -2.6 | ±8.0 | no | 1088 |
| Indian Defense (black) | 93 | -3.5 | ±8.9 | no | 603 |
| English Opening (black) | 83 | +8.1 | ±9.4 | no | 110 |
| King's Indian Attack, with Bf5 (black) | 54 | +7.4 | ±11.6 | no | 133 |
| Scandinavian Defense (white) | 50 | -9.4 | ±12.1 | no | 83 |
| Pirc Defense (white) | 45 | -6.7 | ±12.7 | no | 160 |
Games needed is how many games in that line it would take for an effect of the size you are already showing to become statistically visible. Where it is close to the games you have, the line is worth continuing to watch rather than acting on.
Every one of these 12 lines shows an effect smaller than this sample can resolve. That is the real reading of the empty recommendation list above: not that your repertoire is uniform, but that at 2,627 games spread across 12 lines, nothing you are doing is far enough from par to separate from chance. The honest summary is no fixable leak is visible yet, which is not the same claim as no leak exists.
Each row replays your games in one opening family and splits them on whether a thematic idea actually appeared. 36 plans were tested, so the raw p-values cannot be read at face value: at p < 0.05 you would expect about 2 to look significant by chance alone. The q column corrects for that across the whole set, and 0 survive it.
| Plan | Games in family | Used in | Worth | p | q |
|---|---|---|---|---|---|
| kingside expansion f4–f5 (Sicilian Defense, white) | 484 | 14% | -17.6 | 0.003 | 0.11 |
| the ...f5 break (Sicilian Defense, black) | 622 | 24% | -8.5 | 0.037 | 0.22 |
| queenside expansion ...b5 (Queen's Indian Defense, black) | 142 | 68% | +18.4 | 0.028 | 0.22 |
| kingside fianchetto ...g6 (Queen's Indian Defense, black) | 142 | 66% | +17.1 | 0.037 | 0.22 |
| the ...f6 break (Nimzo-Indian Defense, black) | 114 | 31% | -23.0 | 0.013 | 0.22 |
| queenside expansion ...b5 (English Opening, black) | 83 | 58% | -18.3 | 0.033 | 0.22 |
| f3 (Sämisch / big centre) (Ruy Lopez, white) | 248 | 21% | -11.6 | 0.074 | 0.38 |
| queenside expansion ...b5 (Sicilian Defense, black) | 622 | 59% | +5.5 | 0.126 | 0.57 |
| problem bishop out before ...e6 (Sicilian Defense, black) | 622 | 47% | -4.9 | 0.164 | 0.57 |
| queens off early (by move 15) (Sicilian Defense, black) | 622 | 7% | -8.7 | 0.188 | 0.57 |
| kingside pawn storm ...g5/...h5 (Sicilian Defense, black) | 622 | 38% | -4.8 | 0.181 | 0.57 |
| castled queenside (Sicilian Defense, black) | 622 | 5% | -9.8 | 0.205 | 0.57 |
| queenside expansion ...b5 (Indian Defense, black) | 93 | 50% | +12.1 | 0.203 | 0.57 |
| the ...f6 break (Sicilian Defense, black) | 622 | 19% | -5.4 | 0.226 | 0.58 |
| the ...e5 break (Queen's Indian Defense, black) | 142 | 32% | -9.3 | 0.274 | 0.63 |
| kingside pawn storm ...g5/...h5 (Indian Defense, black) | 93 | 40% | -10.6 | 0.280 | 0.63 |
| kingside fianchetto ...g6 (Indian Defense, black) | 93 | 60% | +9.9 | 0.321 | 0.68 |
| f3 (Sämisch / big centre) (Sicilian Defense, white) | 484 | 20% | -3.9 | 0.409 | 0.82 |
| kingside pawn storm ...g5/...h5 (Nimzo-Indian Defense, black) | 114 | 42% | -5.5 | 0.523 | 0.99 |
| minority attack (b4 and b5) (Sicilian Defense, white) | 484 | 7% | -0.6 | 0.936 | 1.00 |
Worth is the difference in edge between the games where the idea appeared and the games where it did not, inside the same opening family. Only rows with a bold q below 0.05 survive correction; the rest are shown so you can see the full set that was searched rather than only its winners.
Nothing here survives correction. The strongest raw p-values in this table are what searching fifty splits produces on noise, and acting on them would be acting on chance.
Stockfish 18 evaluated every position in 2,614 of these games at a fixed 150,000-node budget — 107,030 of your own moves. Nodes rather than depth or time, so the numbers do not move with machine load.
Mean centipawn loss 29, mean move accuracy 92.6, blunders 2.8 per 100 moves.
| Phase | Moves | Mean cp loss | Accuracy | Blunders / 100 |
|---|---|---|---|---|
| Opening | 26,019 | 9 | 96.7 | 0.1 |
| Middlegame | 45,673 | 33 | 90.8 | 2.9 |
| Endgame | 35,338 | 38 | 92.0 | 4.7 |
| Time left | Moves | Accuracy | Blunders / 100 |
|---|---|---|---|
| under 5s | 1,058 | 90.7 | 9.8 |
| 5-10s | 1,441 | 91.2 | 8.9 |
| 10-20s | 3,703 | 91.4 | 7.0 |
| 20-40s | 8,501 | 91.1 | 6.5 |
| 40-80s | 17,866 | 90.7 | 5.0 |
| 80s+ | 74,397 | 93.4 | 1.4 |
Time remaining when the move was played. This is descriptive, not causal: hard positions eat clock and also cause mistakes, so some of this gap is the position rather than the hurry.
| Moves | Mean cp loss | Accuracy | Blunders / 100 | |
|---|---|---|---|---|
| You | 107,030 | 29 | 92.6 | 2.83 |
| Your opponents | 106,757 | 33 | 91.7 | 3.35 |
The same engine, the same node budget, the same positions — the cleanest like-for-like comparison available, because you and your opponent faced literally the same games.
| Pieces | Moves | Mean cp loss | Accuracy | Blunders / 100 |
|---|---|---|---|---|
| 1-4 | 2,297 | 21 | 97.2 | 3.61 |
| 5-6 | 2,896 | 36 | 94.1 | 5.42 |
| 7-8 | 4,176 | 38 | 92.9 | 5.51 |
| 9-10 | 5,287 | 42 | 91.6 | 5.60 |
| 11-12 | 5,690 | 40 | 91.5 | 5.17 |
| 13-14 | 6,687 | 42 | 90.7 | 5.16 |
| 15+ | 79,986 | 25 | 92.7 | 2.03 |
Every ply tagged with the actual piece count rather than a phase label. The phase boundary is a rule of thumb; this is the quantity the question is really about, and it separates the same games more cleanly.
| Moves later | Blunder rate | Rate otherwise | Difference | 95% CI | p |
|---|---|---|---|---|---|
| 1 | 11.1 | 2.7 | +8.4 | +7.2 to +9.7 | < 0.001 |
| 2 | 9.7 | 2.8 | +6.9 | +5.7 to +8.1 | < 0.001 |
| 3 | 7.8 | 2.9 | +4.9 | +3.8 to +6.0 | < 0.001 |
| 4 | 9.0 | 3.0 | +6.0 | +4.7 to +7.3 | < 0.001 |
| 5 | 7.5 | 3.1 | +4.4 | +3.3 to +5.5 | < 0.001 |
| 6 | 8.0 | 3.2 | +4.8 | +3.5 to +6.1 | < 0.001 |
Rates pool every move, which is the honest estimate — averaging per-game rates would bias it upward, because a short game with one blunder reads as a high rate and then counts the same as a long one. The interval instead comes from resampling whole games, so the fact that moves inside a game are not independent is priced into the uncertainty rather than ignored. Testing per move would produce far smaller p-values here, and they would be wrong.
You reached a position the engine scores at +3.00 or better in 1,734 games (66% of the analysed set) and scored 85.0% from them (95% CI 83.2–86.6). 318 of those games were not won.
No Elo figure is attached to this. Converting every winning position is not available to anyone, so the gap to 100% is not a gain you could bank, and quoting it as one would be inventing a number.
Every position in the analysed games was checked for a tactical opportunity, and the move you actually played was compared against it. 22,091 tactical moments across 2,509 games; you found 82.2% of them overall.
| Motif | Moments | Found | 95% CI |
|---|---|---|---|
| forced mate | 386 | 68.9% | 64.0–73.8 |
| quiet / positional | 3,993 | 74.4% | 72.9–75.9 |
| check / king attack | 2,045 | 75.3% | 73.3–77.5 |
| wins a loose piece | 2,495 | 76.8% | 75.1–78.4 |
| fork | 1,238 | 76.8% | 74.3–79.3 |
| back rank | 970 | 78.6% | 75.8–81.1 |
| discovered attack | 571 | 80.6% | 77.3–83.9 |
| skewer | 1,977 | 81.0% | 79.4–82.7 |
| passed pawn push | 634 | 81.2% | 78.1–84.6 |
| pin | 408 | 81.4% | 77.5–85.0 |
| wins a pawn | 1,130 | 87.5% | 85.5–89.6 |
| favourable exchange | 815 | 90.1% | 87.6–92.4 |
| removes the defender | 2,293 | 91.3% | 90.2–92.4 |
| hanging piece | 3,078 | 96.6% | 95.9–97.2 |
Sorted worst first. Red is below your own overall rate of 82.2% with the interval clear of it, green above. Intervals come from resampling whole games, because several tactical moments in one game are not independent observations.
| Time on the move | Moments | Found |
|---|---|---|
| under 1s | 8,688 | 90.0% |
| 1-2s | 6,595 | 86.8% |
| 2-4s | 3,374 | 75.8% |
| 4-8s | 2,207 | 64.2% |
| 8s or more | 1,227 | 51.8% |
This runs the wrong way at first glance — the longer you thought, the less often you found it — and it is not evidence that thinking hurts. The causation runs the other way: you stop to think precisely when the position is hard, so this is a map of difficulty rather than of effort. What it does say is that the positions where you spend your clock are ones you get wrong roughly half the time — which is where study would actually land.
Each of these is a position from your own games where the engine had you at +3.00 or better and the move you played threw most of it away. They are sorted by how much was lost. 5 of 20 had a quiet winning move — no capture, no check — which is the pattern worth noticing: the wins that get missed are the ones that do not announce themselves.
Engine had you at +5.3. You played Kh2; Kh1 wins (Kh1 gxf3 Re1 f2 Rf1 Kg4). That dropped 1532 centipawns — a quiet winning move.
2026-07-30 · opponent 2542 · endgame · 26s on the clock · game
Engine had you at +10.0. You played Qxd8; Qh4 wins (Qh4 Rg8+ Kh6 Rxg6+ hxg6 h3). That dropped 1359 centipawns — a quiet winning move.
2026-07-13 · opponent 2581 · middlegame · 103s on the clock · game
Engine had you at +3.3. You played Ba7; Kd7 wins (Kd7 e8=Q+ Kxe8 Kc4 Kd7 Kb5). That dropped 1326 centipawns — a quiet winning move.
2026-06-09 · opponent 2470 · endgame · 13s on the clock · game
Engine had you at +2.7. You played Rb1; Bg7 wins (Bg7 Qf2 Kg8 Qg3 Bg4 Qf2). That dropped 1267 centipawns — a quiet winning move.
2026-07-17 · opponent 2587 · middlegame · 0s on the clock · game
Engine had you at +4.2. You played Qb5; Nde5 wins (Nde5 Qb3 Nf3+ Kd1 Bxe6 Nxe7+). That dropped 779 centipawns — a quiet winning move.
2026-07-17 · opponent 2586 · middlegame · 104s on the clock · game
Engine had you at +10.0. You played Qh7; Qxh8 wins (Qxh8 d4 Qc8#). That dropped 1539 centipawns — a forcing winning move.
2026-06-18 · opponent 2507 · endgame · 0s on the clock · game
Engine had you at +10.0. You played Kf3; Rh8# wins (Rh8#). That dropped 1482 centipawns — a forcing winning move.
2026-06-04 · opponent 2583 · endgame · 4s on the clock · game
Engine had you at +4.6. You played Rac8; Rxe5 wins (Rxe5). That dropped 1462 centipawns — a forcing winning move.
2026-07-22 · opponent 2517 · middlegame · 44s on the clock · game
| Situation | Games | Edge here | Edge otherwise | Difference | p | |
|---|---|---|---|---|---|---|
| Right after a loss | 765 | +2.7 | +0.2 | +2.4 | 0.189 | not significant |
| Right after a win | 1378 | +0.4 | +1.5 | -1.0 | 0.544 | not significant |
| Eleventh game of a session or later | 1128 | +0.1 | +1.6 | -1.4 | 0.406 | not significant |
Each row compares that situation against every other game in the window, with the game as the unit of analysis. A difference that is not significant is reported as such rather than dropped, because the absence of a tilt effect is as useful to know as its presence.
You finish won games with a median of 58s left and lost games with 37s. 28 games (1.1%) ended in a loss on time.
Source: the complete Lichess export for Meatloaf927, 2,627 rated blitz games from 2026-05-23 to 2026-08-31, restricted to the the whole archive, 100 days.
Every p-value uses the game as the unit of analysis. This is not a detail: testing at the level of moves treats the forty moves of one game as forty independent observations, which they are not, and inflates significance to the point where almost anything looks real. Two-sided Student's t throughout; Welch's variant where two groups of games are compared. Opening tables carry a Benjamini-Hochberg false-discovery rate because testing thirty lines at p < 0.05 produces roughly one and a half false positives by construction.
Sections appear only where the data supports them. Nothing on this page is hand-written about this player: every number, every opening name and every recommendation is derived from the archive at build time.