Trust, in plain words
How the coach is checked
The AI coach does not judge positions by itself. When you ask it to analyse a position or review a game, a chess engine does the judging first, on your device, and the coach is told to treat those numbers as ground truth and explain them. Under every such reply you see an Engine check panel with the exact facts the coach was given, so you can compare.
The engine
- Engine
- Stockfish 18 — the stockfish.js “lite” build (single-threaded WebAssembly, smaller neural network). Open source, GPL.
- Where it runs
- On your phone or computer, inside the app or browser. The engine work happens there; only the result (the position or game moves and the engine’s numbers) is sent to the coach with your question.
- Analyse a position
- Searches to depth 16. The panel shows the depth actually reached.
- Review a game
- Checks every position for a fixed time: 180 ms per move, or 120 ms per move for games longer than 60 half-moves, so a full game finishes in seconds on a phone. Moves that lose 0.9 pawns or more are flagged; the four biggest go to the coach.
The lite build is weaker than full desktop Stockfish, and a fraction of a second per move is a quick search. It is far stronger than any human and reliable for spotting real mistakes, but small differences (say 0.2 pawns) between two moves can change with a longer search.
What is engine-verified
- The evaluation of a position you paste (who is better, and by how much, in pawns).
- The engine's best line from that position.
- In a game review: which moves lost 0.9 pawns or more, how much each lost, and what the engine preferred instead.
What is not checked (yet)
- The explanation in words — the “why”. It is written by an AI language model (Anthropic’s Claude). Nothing compares that prose to the engine automatically, so a sentence can be wrong even when the numbers next to it are right.
- Any move or variation the coach mentions that is not in the engine line shown in the panel.
- Replies in ordinary chat, where you did not paste a position or a game. These carry a “Not engine-checked” label.
We have not yet published an automated test that replays every move the coach cites against Stockfish. Until we do, treat the engine panel as the source of truth and the explanation as a strong club player's commentary: usually helpful, sometimes wrong.
The weakness scan
The scan uses the same Stockfish build on your device at a short fixed time per move, and re-checks a suspected blunder with a longer search before it counts it. When the coach reads your scan, it gets those numbers (average loss by phase, recurring mistake types, openings you lose with) as data.
Try it
Open the coach, tap ♟ Analyze position, paste a FEN, and read the reply next to its engine check.
Open the coach