Chess Pangolin

Almost every chess app now ships an AI coach, and almost all of them share a failure mode: asked why a move was bad, they restate what the engine's line wins and call that a reason. Chess Pangolin was built to refuse that. Its coach is not a language model at all - it is a set of deterministic checks run against the board and the engine's principal variation (a piece left where it can be taken, a threat already on the board, a fork, a discovered attack, a mate threat, a trapped piece, an overloaded defender, a piece chased with tempo). Each sentence is generated by a check that fired, and a test replays that check on the board before the sentence ships. When nothing fires, the app prints that it has not isolated the cause.

That sounds like a limitation dressed as a virtue, and partly it is - but the publisher has done the unusual thing of measuring it in public, and then the rarer thing of unmeasuring it. The /honesty page keeps a frozen set: fifty CC0 rapid games from the Lichess database, 800-1500, the losing side's costliest moves marked for 194 turning points. Until September it graded what the coach said about every one - 103 explained clearly, 51 abstentions, 0 wrong. The 20 September snapshot, taken after a rebuild of the review panel, throws that scoreboard out. It reports what it can still count (115 of the 194 positions carry a primary teaching demonstration, 78 have no isolated cause, 1 is withheld because the evidence does not match), states plainly that the old grade was an older build's result and will not be carried forward, and drops "zero wrong" as a product-wide claim on the grounds that passing a finite set of checks cannot establish it. Publishers do not usually delete their own best number. This one did, and said why.

The product around that review is more conventional and better built than it needs to be. Import runs through PGN or a Lichess username with no account - Chess.com players export the file themselves, there is no username field for them - and reviewed games stay in the browser, and optional sync deliberately never receives the PGN, the opponent, the username or the game URL. The library is 18 courses banded by rating from Boardcraft Foundations (0-800) to Prophylaxis & Restraint (1900-2500), alongside 76,012 real-game puzzles, an openings trainer, the Geometry Gym's machine-checked sight drills, endgame work checked against the Lichess tablebase, and a Mistakes Bin that schedules your own errors back at you. The Pit plays a full game with the coach commenting move by move, or a rated one against bots.

The honest caveats are the shape of the design. A checker-based coach explains what it has checks for, and on its own frozen set it comes up with no cause at 78 of 194 turning points - two in five. Not all of those are strategic; the publisher is careful to say the audit never classified them, and this review will not either. But whatever the reason, two in five is where the player is told that nothing was found, and nothing found is not the same as help. It is web-only and English-only, the company is months old, and the score that made this page worth writing has been withdrawn with no replacement, which leaves a reader trusting the method rather than a number. Against that, the September change runs the other way on price and access: the review, its coaching, the engine analysis, the puzzles and the games against bots are now unlimited without an account, and All Access fell to £2.99 a month. For a player who has been told "this loses a rook for a bishop" one too many times and wants to know why, this is still the most intellectually honest review tool in the directory - and testing that claim now costs nothing at all.