EloCoach Experiments

Real games. Real engines. AI coaching tested against the board.

We ran EloCoach on the recent rated blitz and rapid games of titled players and streamers, on Chess.com and Lichess, and published the coach’s reviews exactly as the app wrote them. Every engine number sits beside the claim it supports. If you find something wrong, tell us and we’ll publish the correction.

How the reviews were made

  • Stockfish 19 at depth 12 analysed every move, on a phone, the way the app does.
  • Maia-3, a model trained to play like humans at a given rating, estimated how hard each move was to find. Above 2600 it reads moves as a 2600 player would.
  • One coach call per game wrote the review. Two coaches with different methods, EloCoach and Apollo, reviewed every game.
  • A script checked every “piece on square” claim against the positions in the game, and we read every flagged claim by hand.

The run, in numbers

Games
162
117 Chess.com · 45 Lichess
Coach reviews
324
2 coaches
Engine time per game
18.7 s
median, on the phone
Review time
22.1 s
median, EloCoach (Apollo 30.1 s)
Board claims checked
2,001
Real errors found
4

The errors we found

  • A fork described as hitting "the king on h1" when the king was on g2.
  • A fork described as hitting "the king on h1" when the king was on g1.
  • A knight on e5 said to defend c3, which a knight on e5 can never reach.
  • A reference to a pawn on e6 that was never there.

Those games are not featured below. Elite players on Lichess play almost only bullet, so few of their games met the blitz-or-rapid rule.

Featured reviews

We feature 9 of the 162 games, chosen for variety: wins and losses, tactics and endgames, both coaches and both sites. Every number above covers all 162.

Are you one of these players? We’ll remove your games on request, no questions asked: [email protected].