♚Open in the app

Methodology

Puzzles come from two sources: positions mined from the project's own engine self-play games in all 960 starting setups, and — for classical chess — the open Lichess puzzle database (CC0).

How every puzzle is verified

Each candidate is replayed by Stockfish at depth 20 or deeper: the solution line is played out against the engine's best replies, and the position is rejected unless the winning idea is forced.

Puzzles that pass get a verified mark, an engine evaluation with search depth, a human-readable name and a link to their starting position. Rejected candidates never ship.

How strong the computer opponent is

Next to every board you choose the opponent’s level, from about 800 to 2000. From 1400 up it is Stockfish’s own strength limiter (UCI_Elo). Below that the limiter does not work: Stockfish silently raises anything under 1320 to its minimum, so the weaker levels are made another way: less time to think, and at the very bottom a lower skill level.

The numbers for those weaker levels are measured, not assigned: each level plays matches against a reference opponent, and its rating is read from the results. It is the same ladder the mentors use in the app, so “1200” on the site and “1200” in the app are the same opponent. Treat the numbers as approximate.

Open data · Nomenclature