Repository navigation
v1.1.31
·
279 commits
to main
since this release
[1.1.31] - 2025-12-24
Added
- "Max Thinking" difficulty with time-based iterative deepening (capped depth).
- Human vs AI "Play as" selection (White or Black, default White).
Changed
- Play-for-Win now applies a fairness window to avoid large eval drops when avoiding repetition.
- Add timing and symmetry tests for the Play-for-Win guardrails.
Fixed
- Force the default camera to the White view on launch.
- Correct board coordinate label mapping for White/Black views.