Game diagnosis

What this diagnosis does

Enter your chess.com or Lichess username. Your recent games are analyzed with Stockfish and the results combined, so you can find the tactics you keep missing or allowing and review those positions.

The analysis runs in your browser and needs no sign-in. Your browser fetches the games from the platform itself, the engine runs in a Web Worker, and the report is written to your browser's own storage. If you are signed out, that is the only place it goes: clear that storage and the report is gone. If you are signed in, the coaching progress you start from the report is also saved to your account: the tactic, the cause you picked, the review schedule, your chess.com or Lichess username, your practice results, and up to two example positions per tactic with the game link and the move you played. You can delete that progress from the coaching panel, and deleting your account removes it too.

Which games it uses

Rated games in the time control you pick, from your last 50. You can choose rapid, blitz, bullet or classical, one at a time. They are not mixed because time pressure changes when you first fall behind, so a combined figure would describe none of them.

The peer comparison, though, has four conditions: rapid, depth 12, Lichess games, and a rating between 800 and 2199. Its baseline was measured on Lichess rapid games at depth 12, across that rating range only. Comparing anything else against it would mean measuring with the wrong ruler: you would read the gap between rulers, not between players. So that panel is left out instead of filled with an invented number. The tactics table and the first-behind move come from your own games alone, so they appear either way.

Games come only from your latest 50; older games are not fetched. Newly analyzed games are added to your existing results, and up to 500 games are kept.

How a move is judged

Positions where it was your turn are evaluated at your selected depth: 12, 16 or 20, with 12 as the default. The peer comparison is available only at depth 12, where its baseline was measured. A deeper search may change the evaluation or suggested move and takes more time.

The report uses the categories below. Missed, excluded and allowed are distinguished by a 100-centipawn (cp) drop in engine evaluation across your move. 100cp is 1.0 on the evaluation scale; it does not mean you actually lost a pawn.

For example, if a fork was available but you played another move with an evaluation drop of less than 100cp, that opportunity is excluded from the discovery rate. This does not necessarily make your move good; it means the move did not meet this diagnosis's threshold for a miss. Excluded opportunities are counted separately but do not enter the discovery rate's numerator or denominator. A tactic detected in the engine's best reply does not mean your opponent actually played it.

What the report shows

• How many games were analyzed.
• The move number where you are typically first behind, as a median over the games where it happened.
• Losses recorded as timeouts or abandonment, counted separately regardless of the position at the time.
• The same figures split by the color you had.
• Per tactic, what you spotted and what you allowed. The counts and the scenes behind them are shown even from a small sample. Where the sample is large enough, a discovery rate and a per-100-moves frequency are added, because comparing a long game with a short one takes a frequency rather than a count.

What "first behind" does not mean

The move where you first fall 1.5 pawns behind reads like the move that lost the game. We checked whether that is true, on 5,000 rated Lichess games for each time control, using the evaluations Lichess had already published.

Those are the shares that did not go on to lose. Nowhere in the table does the share suddenly drop off. Even after the evaluation reached −12.0, about one in five rapid games did not end in a loss. Pawn units here describe the evaluation, not the actual material balance. The first-behind move can be a starting point for review, but it does not predict the result.

Method, because it changes how the numbers read: only the side that fell behind first is counted, so one game never enters twice. Games lost on time are dropped, which matters most in bullet, where that removed 2,610 of the 5,000 games. A position counts only if at least 20 half-moves followed it, so the sample focuses on positions with time left for a recovery. An evaluation drop on the final move does not itself mean the game was lost.

Two limits come with these numbers. The sample is games Lichess had already evaluated, and analysis there is requested or switched on rather than applied at random, so closely fought games may be overrepresented and the comeback shares may read a little high; that is the weakest point in the measurement. And the deep end of the table rests on a few hundred games rather than thousands, so the 12 pawn row is the roughest of the three.

Why the rates come later

Counts and positions appear even with a small sample. You can review each flagged move from your games. Suggested continuations come from the engine and may not have been played in the game; the tactic labels are automated detections, not a complete account of your mistakes.

Rates are different. A rate from a small sample flips easily with the next few games, so judging your skill or a weakness from it is often wrong. That is why a rate is added only where the sample is large enough.

A discovery rate is added once 97 opportunities have been judged. That number is 1.96² × 0.25 / 0.10², an approximate sample-size target for a 95% interval with a margin of about ±10 percentage points under independent sampling. It is not a guaranteed margin for these game records. It counts opportunities rather than games, because games are not interchangeable: if forks keep turning up in your games, 60 games may be enough, and if they do not, 120 may not be.

The per-100-moves frequency is set by a different formula. It is a rare event, so its precision comes from how many times it happened rather than from how many moves you played, and it is added once 97 occurrences are in. That number is 1.96² / 0.20², an approximate event-count target for a 95% interval with a relative margin of about ±20 percent under a Poisson model. Correlated events within games can make that approximation too optimistic. That it matches the discovery-rate number is a coincidence. Basing it on moves instead would mean reading a frequency off a handful of games, which is exactly how numbers like this mislead people. In the dataset used for that estimate, forks reached the event-count target at about 164 games. The number of games needed for another player can differ.

The denominator is the number of your moves the engine actually evaluated this time. Moves that were fetched but not evaluated are left out. The formula also assumes the events are independent of each other, and scenes that follow one another inside a game are not, so the page shows how many games the scenes came from next to the count. Counts and scenes stay in view before any rate appears, so "measuring" is a state, not a fault.

Questions

Q. Do I need an account on this site? No. The diagnosis works without signing in, and nothing is stored on our server while you are signed out. Signing in adds only one thing: saving your coaching progress to your account. The username you give is a chess.com or Lichess one, so your browser knows whose games to fetch.

Q. Where do my games go? The games themselves are not uploaded. Your browser fetches them from the platform, analyzes them in the tab, and writes the result to your browser's storage. If you are signed in, the coaching progress saved to your account includes up to two example positions per tactic, each with the game link, the position and the move you played.

Q. Why is the peer comparison missing? One of its four conditions is not met: the time control is not rapid, the depth is 16 or 20, the games came from chess.com, or your rating sits outside 800-2199. The baseline was measured on Lichess rapid games at depth 12, within that rating range. A chess.com rating does not mean the same strength at the same number, and the gap has never been measured here, so it cannot be corrected for either. The diagnosis itself still works; only the comparison is left out.

Q. How is this different from the single-game analysis? Game analysis walks through one game move by move on a board. Diagnosis combines many games to show the tactics you keep missing or allowing. In the diagnosis report you can also replay those scenes on a board, then study and practice the idea in corrective training.