Astralbase

Astralbase is an early Rust prototype for chess retrograde analysis.

Overview

Standard endgame tablebases (such as Syzygy or Nalimov) evaluate perfect-play positions down to scalar states: Win, Draw, or Loss. Astralbase currently focuses on the lower-level retrograde machinery needed for that work: legal predecessor generation and queue-based propagation from terminal positions.

The current public API exposes a small RetrogradeEngine with Win, Loss, and Unknown states. CGT canonical forms, persistence, and large-scale generation are roadmap items rather than implemented behavior.

Features

• Inverse Move Generation: Backtrack from terminal chess positions to legal predecessor positions.
• Retrograde Propagation: Propagate scalar win/loss distances through a queue.
• Library API: Use RetrogradeEngine from Rust crates and the partizan Python extension.

Architecture

Astralbase is implemented as a Rust library with a small demo binary. The library owns the retrograde engine so downstream binaries and bindings use one canonical implementation.

Usage

cargo run --release

Dataset shard commands used by the Partizan research harness:

cargo run --quiet -- --non-fixture-composed-domain-shard
cargo run --quiet -- --expanded-non-fixture-composed-domain-shard --rows-per-family 10
cargo run --quiet -- --leakage-clean-non-fixture-composed-domain-shard --rows-per-family 10
cargo run --quiet -- --replay-non-fixture-composed-domain-shard /tmp/astralbase-w22-expanded-composition.jsonl

Research Context

This engine is the core dataset generator for the Partizan research project, aimed at proving that deep reinforcement learning models can learn game-theoretic representations when provided with combinatorial ground-truth data.