[Release] TacTurn V1 — UCI engine with HCE + NNUE residual hybrid

[Umar Khan]

Hello everyone,

I'd like to announce the first public release of TacTurn (V1), a standalone, open-source UCI chess engine written from scratch in C++.

• GitHub:
https://github.com/umarzo/tacturn
• Releases / Downloads: https://github.com/umarzo/tacturn/releases/tag/v1
• Author: Umar Khan (India)
• License: MIT

Key Technical Highlights:

1. Board Representation & Move Generation:
- Standard 64-bit bitboard architecture with runtime magic bitboards.
- PEXT bitboard acceleration enabled on BMI2-capable architectures.

2. Search:
- Iterative deepening Principal Variation Search (PVS) with Lazy-SMP threading.
- Heuristics: Null Move Pruning, Reverse Futility Pruning, Late Move Reductions (LMR), and history-based move ordering.
- Deterministic bench node count across compiled tiers: 188,416 nodes at depth 9.

3. Evaluation (HCE + NNUE Residual Hybrid):
- Base Eval: Hand-crafted evaluation (material balance, piece-square tables, pawn structure, king safety) tuned via Texel's tuning method.
- Neural Net: A small quantized residual NNUE (3072→512×2→32→32→1, king and output buckets, int16 SIMD) trained on 21.5M self-generated positions using Stockfish 17.1 static evaluation labels.
- The residual network acts as a positional correction layer on top of the HCE baseline, yielding +162 ± 17 Elo over pure HCE in internal testing.
- Network weights are embedded directly into the binary—no auxiliary network files are required.

4. Originality & Provenance:
- 100% clean-room source code (MIT license).
- Stockfish was utilized strictly as an external training label teacher and rating sparring partner; no code, weights, or internal structures were copied or derived from Stockfish.
- A full SPRT-gated development history documenting tested heuristics is available in HISTORY.md in the repository.

Binaries Available:
Windows 64-bit and Linux executables (SSE2, AVX2/BMI2, and AVX-512) are statically linked with no external runtime dependencies.

Estimated Strength:
Internal testing against anchored pools at CCRL-like conditions indicates ~3250 Elo.

I welcome any feedback, match testing, bug reports, or discussion on the architecture!

Best regards,
Umar Khan