Rule sensitivity · 22
Can komi or rules flip the AI's best move?
Komi moves the winning threshold; area, territory, ko, suicide, and tax rules change the final ledger. The useful question is not how much win rate moved, but where candidate order actually crosses.
Komi first moves the winning threshold
Varying komi on one fixed board produces a win-probability curve. SAI models multiple komi with a sigmoid, showing that a position does not own one context-free win rate.
Boards A and B are identical except for komi; inspect the curve before hunting a flip. A score shift is not a best-move shift
If every candidate moves together, win probability changes while order=0 remains fixed. An answer flip requires candidate value curves to cross.
The crossing of candidate curves matters more than one win-rate delta. Area and territory price endgame moves differently
Area scoring counts living stones on the board; territory scoring depends on surrounded points and captures. Buttons and seki taxes can also move half-point boundaries and reorder endgame plays.
The same local endgame has a different marginal ledger under area and territory scoring. Ko rules change the legal future
Simple ko, positional superko, and situational superko impose different constraints. A current move can be legal in each while later recaptures and exchanges are priced differently.
Two lines fork at one future recapture: one remains legal and one is closed by superko. Engine and network must both understand the rule
KataGo's rules documentation warns that an old network can mis-evaluate intricate rule situations even inside a new engine. A rule experiment records both identities.
A new engine paired with an old network is not automatically a valid rule comparison. Scan a flip interval instead of choosing two komi
Freeze board, model, visits, and randomization; scan komi in half-point steps and retain the top three moves for each ruleset. Report a crossing interval rather than false precision.
Discrete samples reconstruct candidate ranking bands and the first stable crossing interval. Turn the flip atlas into a teaching condition
Do not call a move eternally best. State the rules, komi range, and position boundary where it is efficient so a player can reassess on another server or tournament.
The final card joins coordinate, applicable rules, komi range, and tradeoff.
Rules and method sources
Before a rule-flip experiment
- Freeze board and model
- Scan komi as an interval
- Separate ranking from win rate
- Record engine and network
- Validate legality under each ruleset