Manaka
CPU-first evaluator line built around incrementally updated sums. The native design avoids shogi-specific hand-crafted cross-features; separate KP/HalfKP experiments exist only as controls for measuring what those features buy.
Manaka's search is PUCT MCTS, not DLManaka's Gumbel search.
Read by topic
Design
- Architecture — current shape grammar and evaluator structure
- Network history — A–E measurements, later aggregation ideas, current switches
- Structural critique — historical diagnosis, reconciled with later measurements
- Search design — placeholder; current PUCT implementation is not yet documented separately
- Inference — placeholder; architecture/benchmark pages currently cover the relevant path
Training and self-play
- Supervised training — controlled distillation experiments; early gen2 results are historical
- Generation loop — ManakaZero self-play → train → gate → external anchor
- Teacher data — shared packed corpus
Measurement
- Manaka benchmark notes — engine-specific interpretation
- Common benchmark — canonical current speed figures
- Throughput work — historical optimization record, not current absolute speed
- Validation results — twelve shapes on one holdout
- KP/HalfKP experiment — feature-space control
Placeholder topics
Value-head decay, self-play plan, and improvement plan are reserved URLs without dedicated Manaka write-ups yet.
The old “loses to material counting” result describes an early v2 policy and is not the current-state summary. Later architecture measurements are the stronger reference.