murk
on crates.io · PyPIA world simulation engine for reinforcement learning: tick-based, deterministic, arena-allocated, with no GC pauses.
Determinism is proved, not assumed — replays hash every tick and report where two runs diverge.
A tick-based core with pluggable spatial backends — 1D lines and rings, square and hex grids, composable product spaces — a propagator pipeline with automatic write-conflict detection, and observation extraction to flat tensors with validity masks, foveation and multi-agent batching. Gymnasium-compatible bindings sit over a batched engine that steps N worlds in one call with a single GIL release. Memory is double-buffered arena allocation; unsafe code is confined to two crates and verified under Miri, and everything else forbids it outright. Deterministic replay records per-tick snapshot hashes and produces divergence reports rather than a pass mark.
cargo add murk
python -m pip install murk