00 · start here
Start here: what OverGrid asks an AI to see.
A short guide to the game, its scoring, and why one three-piece turn can create a serious search problem.
OverGrid / AI Field Notes
Two ways to read the board.
Choose a behind-the-scenes account of the game AI, or a player-facing strategy story drawn from the same concrete problems.
16 notes for Behind the AI.
Behind the AI
For game makers: engineering, evaluation, tooling, and failed experiments.
Open →
01 · search foundation
A small board can still have an enormous turn.
Why OverGrid’s first AI problem was not “which move is best?” but “which legal moves deserve a fair look?”
02 · reply search
More lookahead made the AI worse—until we gave it a smaller job.
The first selective-reply experiments taught OverGrid not to turn every tactical question into a minimax sermon.
03 · early autoresearch
Our first hillclimb had no finish line.
The original autoresearch loop found useful hypotheses—and then taught us why capped snapshots cannot crown a champion.
04 · named AI
Four named AIs walked into the same position. They all made the same move.
The day OverGrid stopped confusing character copy with behavioral identity.
05 · termination
Ending a game is a rule, not a shortcut
Why OverGrid rejected two apparently sensible anti-loop rules before trusting another AI result.
06 · fairness debugging
The home side that won before anyone played.
An identical-agent control exposed a giant seat split in OverGrid’s old arena—and taught us to stop treating a leaderboard as an answer.
07 · fairness reset
When the arena lies, do not add more games
OverGrid archived a large AI dataset after two fairness confounds made its rankings untrustworthy.
08 · reproducibility
Autoresearch needs a contract, not just a hillclimb
The rules OverGrid adopted after a plausible AI lead vanished under more seeds, replies, and horizon.
09 · playtest evidence
A player circled one tile. It changed the AI program.
Why readable game records and public tactical receipts became research infrastructure.
10 · candidate supply
The candidate generator is part of the AI.
Petra’s suspected missed height capture revealed why a good evaluator cannot rescue a move that never enters its menu.
11 · responsive compute
The best move may be outside the search budget
Petra’s missed height tile exposed the real bounded-search problem: evaluating a move is useless if the candidate generator never offers it.
12 · game records
We did not want a replay button. We wanted a game record.
Why OverGrid introduced OGN beside verified replays—and what each format is designed to make possible.
13 · human loop
Autoresearch did not make the breakthrough. Playtests gave it something real to optimize.
How OverGrid playtests made deterministic AI research more useful.
14 · adaptive Ender
Ender had an opponent model. It just never got to use it.
An honest account of a non-promotion, an adaptive search breakthrough, and the difference between activation and strength.
16 · table talk
A good taunt needs the whole table.
Why game AI commentary needs the shared match situation—not hidden information, and not just the bot’s own score delta.