Catalyst v3.1.0

Catalyst v3.1.0

What's New

Search Fixes

• Fixed a null-move pruning regression that had NMP gated to cut-nodes only instead of all non-PV nodes — this was quietly costing a lot of search efficiency
• Removed a direct depth reduction on alpha-raise (confirmed anti-scalar by community testing), relying on standard LMR instead
• Restricted RFP at TT-PV and quiet TT nodes for more accurate pruning
• Added repetition-avoidance logic so the engine prefers a non-repeating alternative when clearly ahead

Move Ordering & History

• Split history bonus/malus constants per table type (butterfly, piece-to, pawn, continuation) instead of sharing one scale
• Skip history updates on trivial single-move cutoffs at low depth to reduce ordering noise
• Restored plain gravity-based updates for continuation history

Transposition Table

• Switched to a 10-byte entry, 3-per-cluster, 32-byte layout — 50% more TT capacity for the same hash size
• Tuned age-weighting in the replacement policy for more accurate eviction of stale entries
• Allocate TT with MAP_HUGETLB, falling back to transparent huge pages (THP)

General Fixes

• Removed mid-search periodic info prints
• Fixed best_thread() score priority logic
• Various stability and correctness improvements

Performance

Elo  | +35.74 ± 6.80 (95%)
Conf | 10+0.10s | Threads=1 | Hash=128MB
Games| N: 4000 | W: 1065 | L: 655 | D: 2280
Penta| [42, 366, 859, 606, 127]
LOS  | 100.00%

Downloads

Pick the binary for your CPU:

Binary	Requirements
Catalyst-v3.1.0-avx512vnni	AVX-512 VNNI (Zen 4, Cascade Lake)
Catalyst-v3.1.0-avx512	AVX-512 + BMI2
Catalyst-v3.1.0-bmi2	AVX2 + BMI2
Catalyst-v3.1.0-avx2	AVX2
Catalyst-v3.1.0-sse41	SSE4.1 (older CPUs)
Catalyst-v3.1.0-x86-64	Any x86-64 CPU

If unsure, use avx2.