Loserfish

A UCI chess engine derived from Stockfish whose goal is the opposite of normal play: to get checkmated. It never wants to win; if every legal move mates, it plays one only because a move is required.

Model

Outcomes for the engine, best to worst: it is checkmated; draw; it checkmates the opponent. By default the opponent is assumed to play the worst moves against that goal (it avoids mating us and tries to force us to mate it). That is misere chess: zero-sum, so alpha-beta applies. Option LoseOpponent=cooperative instead assumes the opponent mates whenever it can. LoseOpponent=adaptive (experimental) searches for a forced loss first, then plays for the opponent type the game so far suggests.

Search (src/misere.cpp): iterative deepening PVS, lock-free shared transposition table, mate distance pruning with misere bounds, LMR, history/killers, free check evasions, Lazy SMP (Threads). Details and status: LOSERFISH.md.

Build

cd src
make -j build ARCH=x86-64 COMP=gcc      # produces src/loserfish (C++17)
# ARCH depends on your CPU: x86-64, x86-64-avx2, armv8, apple-silicon, ... ("make help" lists them)

CMake (no sh needed, GCC/Clang only):

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j

Verified on Linux (perft passes). On Windows (MinGW) it is unverified: use -G "MinGW Makefiles". MSVC is not supported. NNUE and nets were removed; no downloads needed.

UCI options added

Option	Default	Meaning
LoseMode	true	false falls back to the plain Stockfish search (material eval only)
LoseOpponent	adversarial	adversarial, cooperative or adaptive
LoseScore	neutral	score shown to GUIs: neutral (always cp 0), natural, inverted. Neutral avoids GUI resign adjudication
LoseExtCap	3	adversarial search only: free check evasions allowed per path (0-99). Mutual checks otherwise never consume depth; 99 disables the cap
LoseSafetyDepth / LoseMateDepth	6 / 8	cooperative mode only
Threads, Hash	1, 16	Lazy SMP threads and table size
Move Overhead	30	ms taken off movetime and off the clock for GUI latency; movestogo is honored (budget = time / movestogo)

Speed

make profile-build ARCH=<your arch> COMP=gcc (PGO, uses bench, default depth 8) gave about 23% more nps than the plain build in one x86-64 comparison (3.13 to 3.85 Mnps, same node counts). tests/build_variants.sh builds ARCH variants plus PGO; tests/bench.py compares binaries and measures Lazy SMP scaling (time to fixed depth). SMP scaling has not been measured on a multi-core machine yet.

Tests

tests/perft.sh (node counts), tests/lose_selftest.cpp (optimized search equals plain minimax), tests/loserfish_play.py (play-outs, needs python-chess). See tests/README.md. CI: .github/workflows/ci.yml.

Limits (honest)

It cannot force a loss against an opponent who refuses to mate, except through zugzwang-like lines; two adversarial Loserfish engines draw by repetition. Evaluation is simple and untuned. Ponder, infinite, searchmoves, Chess960 and 32-bit builds are untested. Multi-thread speedup has not been measured yet; thread safety was checked with ThreadSanitizer.

WebAssembly

Live demo: https://berat121.github.io/Loserfish/ (source in docs/index.html, docs/app.js; vendored chess.js and coi-serviceworker under docs/).

Prebuilt browser files (about 520 KB .wasm, Emscripten 4.0.9): docs/loserfish.js and docs/loserfish.wasm. Load the script, create the engine with const M = await Loserfish({print: line => ...}), then send UCI lines with M.ccall('loserfish_cmd', null, ['string'], ['go depth 8']). Engine output arrives in the print callback.

• The engine runs on a pthread, so the page needs cross-origin isolation (COOP/COEP headers) for SharedArrayBuffer.
• Rebuild with scripts/build_wasm.sh web|node [outdir] (Emscripten must be on PATH).
• node tests/wasm_smoke.cjs build-wasm/loserfish.js checks the Node build: depth 8 gives the same 612803 nodes as the native binary. The browser build is not tested in a browser yet.

License

GNU GPL v3 (Copying.txt). Derived from Stockfish; see AUTHORS for the Stockfish developers.