ManakaZero generation loop

ManakaZero is the self-play line: gen0 starts from random weights, and each generation produces games, trains a candidate, measures it, and promotes it only when the configured gates pass.

The implementation is in crates/manaka-{selfplay,train,loop,usi}.

One generation

1. Self-play — generate games with the current parent on idle cluster CPU cores.
2. Training — sample from the recent replay window and train the candidate, normally on CUDA.
3. Parent gate — compare candidate and parent with an SPRT.
4. External anchor — compare against a fixed reference so the whole lineage cannot drift unnoticed.
5. Promotion — replace the parent only when the configured checks pass.

The loop writes its ledger under runs/manaka/, so it can resume after interruption.

Data and formats

• weights: MNW1;
• self-play records: MNR1;
• positions are stored packed rather than as architecture-specific features, so old self-play data can be reused after many architecture changes.

The training replay window mixes recent generations instead of using only the latest games.

Search

Manaka uses PUCT search. The normal playing/search budget and the self-play budget are separate settings; do not infer one from the other. Mate-in-one and repetition handling are rule-side fast paths.

Current search throughput belongs in the benchmark.

Running

cargo build --release -p manaka-train --features cuda
cargo build --release -p manaka-selfplay -p manaka-usi -p manaka-loop

build/target/release/manaka-loop probe
build/target/release/manaka-loop run
build/target/release/manaka-loop run --local
build/target/release/manaka-loop watch

For static cluster binaries:

cargo build --release --target x86_64-unknown-linux-musl \
  -p manaka-selfplay -p manaka-usi

watch reads the ledger/progress files and shows generation, phase, node utilization, training status and game-measurement progress. It does not control the loop.

Failure handling

A repeated failure to promote is treated as a signal to inspect the architecture or training procedure rather than retry forever. The loop can emit a stall report and stop after the configured number of failed generations.

The external anchor is important: a candidate can beat its parent while becoming worse against a fixed opponent. Promotion and diagnosis should therefore not rely on internal head-to-head alone.

OpenBench integration

Each generation can be registered as ManakaZero @genN; worker-side configuration resolves the suffix to the static ManakaZero engine command and supplies the generation's EvalFile.

Operational details of SPRT, ladders and worker cleanup are documented in openbench.