Tideborn

Design, training and measurement notes for two shogi-engine lines.

	DLManaka	Manaka
evaluator	residual CNN	incrementally updated symmetric/lookup evaluator
main hardware	GPU	CPU inference, GPU training
search	Gumbel AlphaZero MCTS	PUCT MCTS
training	distillation → self-play	supervised experiments + zero-knowledge self-play line
goal	practical DL shogi inference, including Apple silicon	NNUE-class CPU efficiency with less hand-designed structure

Both lines can read the same packed teacher corpus, but they do not use the same evaluator or the same search implementation.

Start here

• DLManaka architecture / Manaka architecture
• Training and training-data formats
• Current performance benchmark
• Playing-strength measurement
• Latest Manaka result records

Historical diagnoses remain useful, but are not current-state summaries: DLManaka's first self-play plateau and the early Manaka structural critique.

Reading measurements

Treat a number as a measurement only when its conditions are stated. For current speed, use the common benchmark; design pages intentionally avoid maintaining duplicate absolute-speed tables.

For strength, prefer agreement between held-out metrics and games against a fixed external reference. Internal child-vs-parent Elo alone can give the wrong conclusion.