Ravager 3

Search (new)

• correction history: pawn, non-pawn, minor-piece, continuation
• threat-aware history + capture history, real pawn-structure history, 1/2/4-ply continuation history
• triple and negative extensions
• ProbCut with adaptive depth, hindsight depth adjustment, internal iterative reductions
• smarter LMR (node type, TT-PV, correction size, cutoff count) with deeper/shallower re-searches
• qsearch: SEE ordering, futility pruning, full check evasions
• cuckoo tables for upcoming repetitions
• time management scales with best-move stability, node share, failed lows, forced moves
• Lazy SMP thread voting

Transposition table

• rebuilt: lockless, 3 entries per 32-byte bucket, stores static eval + PV flag, age-aware replacement, huge pages
• fixed key/bucket bit overlap bug

Speed

• split accumulators: king move refreshes only the piece-square half, cached per bucket
• lazy incremental accumulators from nearest computed ancestor
• neurons in co-activation order so the sparse layer skips more dead groups
• new AVX-512 VNNI kernel, scalar/AVX2/AVX512/VNNI bit-identical

UCI

• ponder actually works now (go ponder + ponderhit)
• go nodes / go mate, MoveOverhead option, Hash 1-65536 MB
• EvalFile loads external nets of the same format
• built-in bench, perft, d

Estimated rating: ~3550 Elo (40/15 time control, over 500 games)

Ravager Fury

Features

• Replaced Leorik NNUE with a new embedded network and optimized C11 inference.
• Added Lazy SMP multithreading with private worker search/evaluation state and a shared, thread-safe transposition table.
• Implemented full Chess960 support, including X-FEN/Shredder-FEN castling rights and overlapping or stationary castling.
• Improved move ordering, pawn history, 4-ply continuation history, SEE checks, LMR, and null-move verification.
• Enhanced time management with root-move node accounting and asynchronous UCI search for responsive stop and isready commands.
• Improved network-load validation and Windows memory alignment.
• Removed the zstd dependency for self-contained Linux and Windows builds.
• Added regression tests for NNUE, Chess960, perft, and multithreading.

Match Score (Ravager Fury vs Ravager 2.1)

Metric	Result
Elo Difference	+161.29 ± 6.56 (95% confidence)
Time Control	60.0+0.60s
Threads	1
Hash	64 MB
Games Played	3,645
Wins	1,814
Losses	236
Draws	1,595
Pentanomial	[6, 40, 386, 1146, 2421]

Builds & Compatibility

• Self-contained Linux and Windows builds.
• Available in ZIP and TAR.GZ archives.
• No external zstd dependency required.

Ravager 2.1

Downloads

file	platform	notes
ravager2.1-linux-x64.tar.gz	Linux x86-64	AVX2/BMI2/POPCNT, glibc
ravager2.1-windows-x64.zip	Windows x64	static build, no DLLs

Both binaries are fully self-contained: the NNUE net is embedded, the texel-tuned HCE is the built-in fallback. Unpack, point your GUI (UCI) at the executable and play. The only optional configuration is SyzygyPath for endgame tablebases — everything else works out of the box.

Strength

• vs Ravager 2: +50 ± 20 Elo over 500 games (40/15 + 0.1s, 250 mirrored pairs)

Improvements in Ravager 2.1

• Search & Move Ordering:
  • Singular extensions with double extension and multi-cut pruning
  • Late Move Reductions (LMR) informed by combined continuation history (CMH)
  • Full SEE-based capture ordering and new capture history heuristic table
  • Quiescence search with TT probing and storing
  • Adaptive null-move pruning and losing-capture pruning
  • Dynamic score-adaptive aspiration windows
• Transposition Table:
  • Cache-line prefetching to reduce memory latency
  • Age-penalized quality replacement strategy
• Evaluation:
  • Rook on 7th rank bonuses
  • King safety contact check evaluation
  • Blocked passed pawn discount and weak square detection
• Time Management:
  • Phase-aware time budgeting with score oscillation protection

Build from source

make            # self-contained binary (~5 MB)
make EVALFILE=  # HCE-only binary (~170 KB)

Ravager 2

Downloads

file	platform	notes
ravager2-linux-x64.tar.gz	Linux x86-64	BMI2/POPCNT, glibc
ravager2-windows-x64.zip	Windows x64	static build, no DLLs

Both binaries are fully self-contained: the NNUE net is embedded, the
texel-tuned HCE is the built-in fallback. Unpack, point your GUI (UCI) at the
executable and play. The only optional configuration is SyzygyPath for
endgame tablebases — everything else works out of the box.

Highlights

• NNUE evaluation — (768→H)x2 dual-perspective accumulator; loads
  Ravager-format and Leorik-format bucketed SCReLU nets natively. Default
  build embeds the 640HL-S-5io8 Leorik net via incbin.
• Syzygy tablebases (Pyrrhic, vendored): WDL probing in search,
  DTZ-optimal root moves. UCI: SyzygyPath, SyzygyProbeLimit,
  Syzygy50MoveRule.
• Texel-tuned HCE: tuner v2 with EPD datasets + golden-section K search;
  parameters retuned on the Zurichess quiet-labeled set.

Measured strength

• vs Ravager 1 (2+0.02, 150 games, LOS 100%): tuned HCE +271 ± 70,
  NNUE +568 ± 178
• Open-engine tournament (10+0.1): 55/60 vs Fruit 2.1, Toga II 1.2.1a,
  Glaurung 2.2, Senpai 1.0, Crafty 25.x, Phalanx XXIII
• BayesElo anchored estimate: ≈2950 Elo

Build from source

make            # self-contained binary (~5 MB)
make EVALFILE=  # HCE-only binary (~170 KB)

See README.md for the full UCI option list and CREDITS.md for bundled
third-party components.

Ravager 1

Ravager 1 — pure handcrafted-evaluation chess engine.

No NNUE. No opening book. No endgame tablebases.

Strength: ~2380 CCRL-blitz scale (95% CI 2320-2445), measured with
cutechess-cli vs CT800 1.46 (CCRL 2708) and Amundsen 0.80 (CCRL 2325),
148 games total, Bayesian Elo fit (Davidson draw model, anchor priors).

Highlights:

• Tapered handcrafted evaluation, texel-tuned (datagen + tuner included)
• PVS search: aspiration, singular + double-check extensions, LMR/LMP/
  futility/razoring, adaptive NMP + verification, ProbCut, improving-scaled
  margins, continuation history, predictive time management
• Incremental PSQT/material/phase accumulators, pawn-structure hash
• BMI2/PEXT sliding attacks (x86-64 with BMI2 required for this binary)
• Verified by exhaustive perft (startpos, Kiwipete, CPW 3-6)

Build from source: make
Full credits: see CREDITS.md