Aether runs a lightweight neural evaluator (feature MLP) plus a move policy prior in the browser. Stockfish 18 WASM powers deep Game Review. There is no cloud GPU on the free site — we do not pretend otherwise.
Feature MLP blends material, mobility, king safety, pawns, center.
Softmax ranking so search looks at good moves first.
Twin bots embed aggression / solidity from repertoire stats.
Shallow neural net evaluates positions; policy prior orders moves for deeper search. Blunder rates still model human rating.
Style embedding from repertoire/capture/check rates clones opponent aggression vs solidity.
Engine + centipawn loss classification (Sigma→Clown); NN eval can re-rank critical moments.
Weakness clustering from stats/streaks — rule + feature hybrid (upgrade path: train on user games).
Elo-ish progression + theme tags; future: difficulty model from solve time / fail rate.
Multi-factor scoring on openings/tilt/clock — ensemble of heuristics ready for supervised weights.
Rating proximity queue; future: Glicko + win-prob model for fairer pairs.