Why We Built an Opening Book with 585,000 Positions

The book is not a list of engine-approved moves. It is a database of positions collected from real human games, annotated with frequency data at different ELO levels.

When you reach a position in the opening phase, the engine checks the book. If the position exists in the database, it returns the most popular move at your ELO level. If the position is not in the book, it falls back to the neural network.

The book covers the first 15 moves of a game. After that, you are on your own. This is intentional. The opening is where knowledge matters most. The middlegame is where understanding matters most. The book handles the first part so the neural network can focus on the second.

This is the part that makes it different from a standard opening book.

A standard book gives you the same move regardless of your rating. If the book says 1.e4 is best, it says 1.e4 whether you are rated 1200 or 2400. This does not match how humans learn.

Our book selects moves based on popularity at your ELO level. A move that is common at 2200 might be rare at 1500. A move that works at 1800 might be a trap at 2400. The book reflects this.

The 585,000 positions were collected from human games. Not engine games. Not self-play. Real games played by real people across a wide range of skill levels.

This matters because engine games produce moves that humans would never play. An engine might find a brilliant novelty on move 8 that wins by force. No human would find that move in a real game. Including it in an opening book teaches you nothing about how humans actually play.

Human games have patterns. Humans favor certain structures. Humans make consistent mistakes in specific openings. The book captures all of this.

Stockfish is designed to find the strongest move in any position. An opening book is a shortcut. It bypasses calculation and returns a pre-computed answer. This is the opposite of what Stockfish is built to do.

Stockfish does not need an opening book because it can calculate the opening from scratch. It has the processing power to evaluate the first 15 moves as accurately as any book. But this accuracy comes at a cost: it tells you what a supercomputer would play, not what a human should play.

An opening book is not about strength. It is about relevance. It gives you moves that are statistically proven to work at your level, against opponents at your level, in positions that humans actually reach.

The book covers 585,000 positions. This sounds like a lot. It is not. The total number of legal chess positions is estimated at 10 to the power of 43. Even restricting to the first 15 moves, there are millions of possible positions. The book covers the most common ones.

If you play an obscure opening or reach a rare position, the book will not have an answer. The engine falls back to the neural network. This is not a failure. It is the expected behavior.

The book is a tool, not a crutch. It gives you a starting point. Understanding why those moves work is still your job.

If you are a beginner learning your first openings, the book gives you statistically sound moves without overwhelming you with theory.

If you are an intermediate player stuck at a rating plateau, the book shows you what stronger players at your level are doing differently in the opening.

If you are an advanced player preparing for a tournament, the book is less useful. You already know your openings. You need deeper analysis, which the neural network provides.

The book is designed for the first two groups. The majority of chess players.