ChessGrader / Methodology

How ChessGrader grades your moves

The public model, thresholds, safeguards, and node budgets behind a ChessGrader report, written down. If you think a grade is wrong, this page is how you check our side of the math.

By the ChessGrader team · Updated July 10, 2026

The engine

ChessGrader runs Stockfish 17.1 with the lite NNUE network, compiled to WebAssembly, in your browser. There is no server-side analysis. Every position in your game is evaluated with a fixed budget of 100,000 nodes per move. The app verifies at startup that the NNUE network actually loaded and records that with each analysis — it never silently falls back to the weaker classical evaluation.

Why fixed nodes instead of depth or time

Most analysis tools search for a fixed time or to a fixed depth. Fixed time means a gaming PC and a five-year-old phone produce different evaluations — and therefore different grades — for the same game. Fixed depth is closer, but depth is an uneven unit: modern Stockfish prunes aggressively, so "depth 20" covers wildly different amounts of work in different positions. A fixed node count makes the search deterministic in effort: the same game gets the same grades on any device, any day. That reproducibility is worth more to a grading system than squeezing out extra strength on fast hardware.

The two-pass design

The first pass grades every move at 100,000 nodes. Then the moments that matter — potential blunders, potential brilliancies, big evaluation swings, mates appearing or disappearing — are re-searched at 600,000 nodes with MultiPV 2 (the engine's two best lines), up to 20 ordinary critical positions per game. Mate-distance changes are always checked, even when that takes the pass over 20.

The point: big labels should rest on deep numbers. Brilliant and Great can only be awarded by this verification pass. A shallow search occasionally mistakes a losing sacrifice for a sound one; six times the nodes and a second principal variation catch most of those. The verification pass also works in your favor on close calls — if your move ties the engine's second line within 10cp in a non-mating position, or keeps the same mate distance, it's credited as Best, compared on one consistent search.

Expected points

Chess.com's February 2026 Classification V2 documentation says ordinary labels are based on rating-aware expected-points loss. The inputs and classification boundaries are public; the fitted conversion model is not. ChessGrader therefore uses a conservative approximation: a player's engine evaluation and rating become an expected score from 0 to 100, and a move is measured by how much of that score it gave up.

Expected points capture the important context: dropping from +2 to 0 matters much more than dropping from +9 to +7. But a probability curve becomes very flat in decided positions, so ChessGrader also keeps a centipawn-loss backstop. That prevents a 100–400cp technical error from being called Good merely because the result was already likely. Centipawn loss explained covers why.

Move grades

There are 11 labels. Ordinary quality labels use the expected-points bands below; Book, Forced, Miss, Great, and Brilliant also need their own chess-specific evidence.

Plain-English versions of each label live in move classifications explained, and the sacrifice test behind the !! label in what makes a move Brilliant.

Centipawn and mate safeguards

Expected-points math needs guardrails at the extremes. Centipawn loss acts as a backstop when the expected-points curve has flattened, but is capped below Blunder in a position whose result was already overwhelming. Mate has explicit rules too: skipping mate in one is a Miss even if another forced mate remains, and losing a forced mate while still winning is also a Miss. Other longer mating routes stay at least Excellent, while a defender already facing forced mate is not charged again for the inevitable.

Accuracy

ChessGrader uses a CAPS2-inspired approximation, calibrated against reviewed games exposed through Chess.com's public game archive. Accuracy uses its own rating-neutral expected-points loss, separate from the centipawn safeguard used for move labels. Book and Forced moves score 100; every other move scores max(0, 100 − 7.2 × loss^1.75), and the game score is the ordinary mean. That school-test average keeps one zero from overwhelming dozens of precise moves. Chess.com's exact CAPS2 formula is private, so benchmark agreement is the calibration target—not pixel-identical output. Benchmarks by rating tell you what the resulting number means.

Estimated game rating

When the player's real rating is available, the game estimate starts there and moves up or down using both expected-points accuracy and capped average centipawn loss. Accuracy measures moves that changed the likely result; centipawn loss catches technically loose play in positions where the result was already nearly decided. The adjustment is capped at 600 points and needs at least 10 moves. Rating-less PGNs fall back to a community ACPL-to-rating fit. Caveats, stated as bluntly as we can: one game is a tiny sample; time control changes everything; a one-sided game can flatter the result. It's a vibe, not a measurement, and the report says so.

How close it actually gets to chess.com

Measured, not asserted. Two numbers matter, and they say different things.

• Accuracy scores: mean absolute error of 4.30 points. Across a corpus of 12 publicly reviewed games and 24 player scores, the median error is 3.06, the largest single error is 14.01, 17 of 24 scores land within five points, and the stronger player is identified correctly in 12 games out of 12.
• Individual move labels: 65% exact agreement. Across 15 games and 406 moves walked out of chess.com’s own Game Review interface, ChessGrader assigns the identical label on 264 of them. Most of the remaining disagreement sits on the Excellent / Good / Best boundary — the region where centipawn loss alone does not separate the classes — rather than on mistakes and blunders, which are the labels that change how you play.

We publish the lower number as readily as the higher one. Any tool claiming near-perfect agreement with chess.com is claiming to have reproduced a private, fitted model; what can honestly be measured is compatibility, and this is ours.

Why your numbers can differ from chess.com's

Chess.com's fitted expected-points model, exact CAPS2 formula, search depths, and Brilliant detection remain private. ChessGrader follows the published Classification V2 concepts and calibrates against public reviewed-game benchmarks, but it does not claim to reproduce those private systems. Borderline grades and whole-game accuracy can differ materially. When the two disagree, you can at least check ChessGrader's side of the argument — it is all on this page.

What the private-beta coach is being built to do

The coach is currently invitation-only and the public route is a waitlist. The dormant beta implementation freezes the latest 30 eligible games, then ranks a repair priority by recurrence across distinct games, impact, trainability, evidence confidence, and the goal the player selected. Every diagnosis carries its denominator, confidence reason, and exact source positions. Results alone cannot create an opening diagnosis; that requires engine-backed evidence across multiple games.

1. Recall: solve cold while the move, evaluation, grade, and old explanation stay hidden.
2. Repair: commit a move, predict the reply, and explain the decision before seeing corrective feedback.
3. Transfer: write a decision rule and apply it to a mixed position or a new matched-time-control game.

The ten-minute schedule is exactly 2 minutes of recall, 5 of repair, and 3 of transfer. Longer schedules preserve the same sequence. Primary positions return on days 1, 3, 7, and 14; mixed positions and unseen game receipts prevent memorizing one board. Missed sessions roll forward without a streak penalty.

The campaign awards practice XP for completed training acts, never for a predicted rating gain. XP persists across missed days, and the player can change their schedule at any time. That choice is deliberate: visible mastery supports competence without using a fragile daily streak as pressure. The approach follows the autonomy, competence, and relatedness principles described by self-determination theory.

The coach waits for ten strictly newer, like-for-like games before retesting the frozen priority. It does not silently switch the focus and it does not claim that a rating change was caused by the plan.