← Devlog ·
A bot for balance, not for fun
What it is
A simulator that plays whole runs with the same logic as the game: same rules, same rewards, same turn manager, no window. About 0.02 seconds per run, so 1,500 runs per policy in about 5 seconds.
What it found
1. Maps with no exit. In roughly 4.4 to 5.8% of maps the only path to the boss went through a room that needs a key you might not have. Fixed with a deterministic retry and tests on every floor with 250 seeds. 2. A boss "impossible" to beat. The simple bot won 0 of 381 fights against the Garden Gnome. 3. A real crash. An enemy leaving the board left its position empty. 3 of 100 games on one floor. Fixed, with a test.
The lesson about bots
The simple bot was the problem, not the Garden Gnome. I built a planning bot that tries the best actions and simulates the next 6 rounds. It won about 9.3 rooms per run, against 3.35. Against the Garden Gnome it won 48%. The Rosebush fell 91% of the time and the Iron Toad 92%. Samples are small. I trust the shape, not the decimals.
What it can't do
It can't tell you whether something is fun. It can flag suspects: most basic enemies die without acting more than 80% of the time against good play. Whether that is "cannon fodder by design" or "boring" is a human call. That's what the beta is for.
Unlocked
Twine Bobbin. Doesn't attack. Leaves an impassable twine trail up to 4 squares long. The numbers say it costs you about 3.3 legal moves per round.
Next
Devlog 004: three bosses in the Garden.