Since you’re here...

We hope you will consider supporting us today. We need your support to continue to exist, because good entries are more and more work time. Every reader contribution, however big or small, is so valuable. Support "Chess Engines Diary" even a small amount– and it only takes a minute. Thank you.
============================== My email: jotes@go2.pl

Simple Chess 3.2.0 JA - new version chess engine


SimpleChessEngine - UCI engine
Author: LinkLikesLattes  Rating Chess Engines Diary CEDR=2269

SCE is a chess engine written in educational purposes. Its rating is aproximatly 1690 - 1750 elo. SCE won shallow blue and tscp It uses UCI protocol to communicate with GUI interface. It was not tested on linux!

What's new?
Chess960 / FRC / DFRC support and a new self-play network, plus an evaluation-tension correction-history update and an endgame search fix.

Strength: +34 Elo over 3.1.0. Head-to-head vs SimpleChess 3.1.0 — 5,526 games at 10s+0.1s, 1 thread, 256 MB hash, paired openings: Elo +34 [+26, +41], LOS 100%. The gain is the network retrain; the code changes are at parity at this fast control and are built to pay off at the longer time controls (2'+1" and up) the search is tuned for. Estimated 3529 on the CCRL Blitz 2'+1" (1-CPU) scale (95% interval 3501–3557) — our own gauntlet estimate, not a CCRL rating.

Correction history weighted by evaluation tension (E6). The network is queried a second time with the side-to-move swapped; the resulting "tension" scales how large a step the pawn-structure correction table takes at deeper nodes. +3.5 Elo at 10s+0.1s.
Adaptive search width no longer narrows in the endgame. The width that scales reductions and pruning used to fall to zero once material came off, so the engine searched its narrowest tree exactly where lines are most forcing; that material factor is gone. +3 Elo, and it fixes the endgame depth-vs-move-stability pathology.
Chess960 / FRC / DFRC. Set UCI_Chess960 (GUIs and lichess-bot do it for you): castling is spoken king-to-rook (e1h1), X-FEN and Shredder-FEN castling fields are both accepted, and asymmetric (DFRC) back ranks need nothing extra. Standard-chess play is bit-identical to 3.1.0. Verified against Stockfish 19's FRC/DFRC perft set; a new perft <depth> debug command is included.
Move Overhead is honoured. The option was parsed but ignored (a hard-coded 30 ms reserve); a GUI or lichess-bot that asks for more is now respected.
Move-generation fix. A vendored library mis-counted double checks from two diagonal sliders, which could produce an illegal interposition or a crash; fixed and verified against python-chess over 1.7M positions. position now also validates every move and leaves the board unchanged on a king-less or unparseable FEN instead of crashing later.
New network. SCNNUEv3-2026-09-12.scn5, trained from the engine's own self-play; the whole +34 Elo is this retrain.
Source release. Build with make (or make profile-build PGO_NET=nets/SCNNUEv3-2026-09-12.scn5 for the fastest binary); the network ships in the repository. Third-party notices are in NOTICE.

Jim Ablett compiles:

Comments