Misty 0.1.3

The belief now records the en passant square under the fog rule instead of standard chess legality. Before this, an en passant capture that left your king attacked dropped the square, so the belief lost the true board: in one production game it went from 625 positions to 41. The fix is in both the Rust and Python paths, with a regression test.

The README adds a Known limitations section linking the strength issues.

Exported from the same engine commit that mistboard.com runs.

Misty 0.1.2

If you installed from PyPI before today, you have the wrong engine. This is the first release that reaches PyPI carrying the profile production runs.

0.1.1 corrected the git tree and stopped there. Its pyproject.toml still declared 0.1.0, nothing was uploaded, and PyPI went on serving the 2026-08-23 build whose newest profile is v1.6-rc1-net-prune, annotated "not shipped". So the defect 0.1.1 was cut to fix survived it on the one channel the README's install line names.

pip install -U misty-chess gets you v1.6, which is v1.5 plus a floor on the catastrophe prune's hang test.

Also fills in the PyPI page. 0.1.0 and 0.1.1 both shipped with no classifiers, no project URLs and no homepage, so the package was effectively unfindable by anything but its exact name.

No engine change. The tree is byte-identical to 0.1.1 apart from pyproject.toml and one .gitignore line. Regenerated at engine dfa6d88, the commit engine.ref pins for production. Gate: leak scan clean, 637 tests collected, and the built wheel installs into a clean 3.12 venv with v1.6 present in PROFILES.

Install is unchanged:

pip install -U misty-chess
brew install stockfish   # or: apt-get install stockfish

The PyPI wheel is still pure Python and roughly 500x slower on the belief hot path; build the Rust extension with maturin develop --release for real strength at real time controls. Binary wheels are tracked in #1.

Misty 0.1.1

Corrects what the 0.1.0 export shipped.

The public tree now contains the profile production actually runs. The 0.1.0 export was cut at 00:13 UTC on 2026-08-23; the v1.6 freeze landed at 03:41 the same day. So the published code topped out at v1.6-rc1-net-prune, annotated "not shipped", and anyone checking which profile serves games on mistboard.com would have concluded v1.5. It is python-v2-v1.6, and has been since the day of the first release.

v1.6 is v1.5 plus a floor on the catastrophe prune's hang test. The prune netted the value of the captured piece, so giving up a queen for a defended rook scored as a 400cp loss rather than a disaster and slipped past the guard. Two production games were lost that way before the floor shipped.

Also removes 186 __pycache__ artifacts (108 tracked files) that rode in on the initial git add -A, despite a .gitignore that excludes them.

Regenerated from engine dfa6d88, the commit engine.ref pins for production. Gate: leak scan clean, 637 tests collected, exported suite 636 passed / 1 xfailed against a fresh virtualenv with the exported fow_rust crate built standalone.

Install is unchanged:

pip install misty-chess
brew install stockfish   # or: apt-get install stockfish

The PyPI wheel is still pure Python and roughly 500x slower on the belief hot path; build the Rust extension with maturin develop --release for real strength at real time controls. Binary wheels are tracked in #1.

Misty 0.1.0

First public release.

Misty is a fog-of-war (dark) chess engine: exact enumeration of every position consistent with what its side has observed, one-sided growing-tree CFR search over sampled information sets, Stockfish depth-1 leaf evaluation, and purification plus a resolve gadget at move commit — the architecture line of Obscuro (Zhang & Sandholm, ICLR 2026). It plays live at mistboard.com.

This release ships the engine core with its Rust-accelerated belief hot path (pure-Python fallback with identical results), a post-game analyzer that classifies mistakes as belief / sample / decision errors, the JSON serving protocol spec, and a 600+ test suite whose parity gates require Rust and Python to agree byte-for-byte across real-game replays.

Install: pip install misty-chess (PyPI upload pending — until then, clone and pip install -e .). Stockfish must be on PATH. See the README for the accelerated build.

License: GPL-3.0-or-later.