Rust Chess Engine — Maximize ELO

Build and iteratively improve a UCI chess engine in Rust. Your engine plays a parallel SPRT (Sequential Probability Ratio Test) gauntlet against Stockfish at calibrated strength levels and CCRL-rated reference engines. The score is your estimated ELO rating — higher is better.

Baseline: ~2435 ELO (ported from github.com/deedy/chess — TT, LMR, null move, PVS, killer moves, SEE, PSTs, king safety) Target: As high as possible (2718 = Deedy's benchmark, 3723 = Stormphrax ceiling) Ceiling: 3723 ELO (Stormphrax, strongest CCRL-rated opponent)

Quick Start

bash prepare.sh          # Install Rust, Stockfish, fastchess, reference engines
bash eval/eval.sh        # Compile engine + run parallel SPRT gauntlet + compute ELO

What You Modify

Everything under engine/ is fair game:

engine/
  Cargo.toml        # Add dependencies (NNUE crates, bitboard libs, etc.)
  src/main.rs       # Search, evaluation, move ordering — the engine itself
  src/*.rs          # Create additional modules as needed

What You Cannot Modify

• eval/ — Evaluation scripts (gauntlet runner, ELO computation)
• prepare.sh — Setup script
• tools/ — Stockfish, fastchess, reference engine binaries

How Evaluation Works

cargo build --release
       |
       v
Parallel gauntlet via fastchess (40 moves / 2 minutes)
       |
       +--> vs Stockfish UCI_LimitStrength (5 levels: 2600-3000)
       +--> vs CCRL-rated engines (Blunder, Inanis, Mantissa, Stormphrax)
       |
       v
SPRT (Sequential Probability Ratio Test)
Terminates early once a result is statistically significant.

Gauntlet Opponents

Opponent	Rating	Source
SF 1000-2000	1000-2000	Stockfish UCI_LimitStrength
Blunder 6.1	2105	CCRL Blitz verified
Blunder 8.5	2667	CCRL Blitz verified
Inanis 1.6	3085	CCRL Blitz verified
Mantissa 3.7	3317	CCRL Blitz verified
Stormphrax 7.0	3723	CCRL Blitz verified
SF depth 4-10	~1500-2700	Cross-validation

Improvement Roadmap

Phase	ELO Range	Key Techniques
Baseline	~2435	TT, LMR, null move, PVS, killer moves, SEE, PSTs, king safety
Core	1500-2000	Transposition table, MVV-LVA, killer moves, null move pruning, LMR
Evaluation	2000-2400	Piece-square tables, pawn structure, king safety, mobility
Advanced	2400-2800	NNUE evaluation, PVS, singular extensions, Syzygy tablebases
Elite	2800+	Multi-threaded search, SPSA tuning, NNUE self-play training

Anti-Cheat

The eval enforces:

• SHA-256 checksums on all tool binaries (no tampering with opponents)
• Source scan for network access, process spawning, protected path reads
• Git diff check on protected files
• Cross-validation between opponent types (detects reward hacking)
• 15,962-position Drawkiller opening book (no memorization)

File Structure

rust_chess_engine/
  program.md           # Full task spec (agent reads this)
  prepare.sh           # One-time setup
  requirements.txt     # Python dependencies
  gen_openings.py      # Opening book generator (fallback)
  engine/              # YOUR CODE — modify freely
    Cargo.toml
    src/main.rs
  eval/                # READ ONLY
    eval.sh            # Compile + gauntlet + scoring
    compute_elo.py     # MLE ELO estimation
    openings.epd       # Fallback opening book (30 positions)
  data/                # Created by prepare.sh
    openings.epd       # High-quality Drawkiller opening book (15,962 positions)
  tools/               # Created by prepare.sh
    stockfish
    fastchess
    blunder-6, blunder-8, inanis, mantissa, stormphrax