What the maps mean
Territory / likely ownership. The local KataGo network predicts final ownership at each intersection. Teal means Black. Coral means White. Low intensity means a weak prediction. These estimates are not a proof of life, death, or secure territory.
Influence. A deterministic field spreads outward from each color’s stones. Opposing stones block it. The field decays across empty intersections. Contours show equal levels of influence. This is a teaching heuristic, not KataGo output.
Moyo. Dashed contours enclose empty points with strong influence from one side. This is a simple framework heuristic. A marked region can still be invaded. It is not a territory count or a tactical verdict.
Groups and liberties. Connected stones form a group. Its liberties are adjacent empty intersections. These counts come directly from the board rules.
Candidate moves. The letters rank legal moves by the network’s policy output. An unmarked move is not necessarily bad.
Play and review
Tap an empty point to play. White replies automatically. Set the opponent and strength in the menu. Undo takes back your move and the reply.
Strength. Balanced compares five candidate moves. Strong also checks three replies to each. Gentle selects a weaker candidate within three estimated points of its best choice. These settings are relative. They are not calibrated Go ranks.
Blunder alerts. Review compares your move with five candidate moves and checks two replies for each. An alert appears when the estimated point loss exceeds your menu threshold. Low move preference alone does not trigger an alert. The search is shallow, so alerts say “Possible blunder.” Review strength stays the same at every opponent setting.
The engine
This app runs a pinned KataGo evaluator in a background worker. It uses single-threaded CPU WebAssembly with SIMD. No remote analysis or LLM service receives your positions. The app ships the small g170-b6c96 model from KataGo’s official January 2020 run. It favors small downloads and quick feedback over the strength of current large models.
The board accepts every integer size from 2×2 through 19×19. The network was primarily trained for normal Go board sizes. Very small and unusual boards can produce weak analysis. The rules still apply exactly.
The evaluator uses a fixed orientation, with no sampling and no time-based search budget. Repeated evaluations are tested for exact equality in the same runtime. Cross-device bitwise equality is not promised.
Rules and storage
Area scoring, no suicide, positional superko, and adjustable komi. Two passes end play. The final count assumes every remaining stone is alive. Resolve disputed dead groups through play before passing. The default komi is 6.5.
Study files preserve every branch. SGF export also includes branches. The browser stores your current study on this device. Download a study file before clearing browser data.
Open source
KataGo and the pinned browser port source provide the engine foundation. The app builds a small CPU evaluator with its own history validation and output handling.