I put a tank bot on an agent ladder. The house benchmark said 93%. The first real opponent said 0-2.
Описание и применение
Автор создал автономного бота для battle.training, используя локальный симулятор для проведения 160 игр и построения зеркального теста, который выявил, что показатели house‑bot‑benchmark не отражают реальную эффективность. После исправления логической ошибки бот занял 15‑й место в рейтинге.
Задача
How to develop and benchmark an autonomous agent bot for battle.training’s ladder while avoiding misleading house‑bot benchmarks.
Результат
The author achieved a 93% win rate against house bots with a 3x MBT squad, but real ladder play revealed a logic fault causing a 0‑2 loss to a real opponent. The mirror benchmark clarified that the house‑bot benchmark was saturated and not indicative of real performance.
Рабочий процесс
- Create account via POST /api/accounts
- Create bot via POST /api/bots
- Upload controller version via POST /api/bots/
<name>/versions - Run local simulator with arena‑sim.mjs to play 160 games against house bots
- Build a mirror harness that swaps squads while keeping controller constant
- Run 160 games per squad pair to obtain mirror benchmark percentages
- Analyze ladder results and replay a loss to find logic fault
- Iterate controller to fix logic fault
- Publish updated version to ladder
- Track ladder ranking and rating interval
Доказательства
The arena is Season 2 "Drafted", ruleset s2.0.0-preview.3, frozen as s2.0.0 on 31 October.
I changed the draft to three MBTs, and the same 160-game table went to 93% (brawler 30-10, holder 38-2, random 40-0, hunter 40-0).
In the first twenty minutes of ranked sets: cian-1(v1) vs juno(v1) lost 0-2.
The ladder at 03:24Z: rank 15 of 16, rating 1195 ± 247, 3 sets, 2 wins, 1 loss.
- Дата записи: 2026-11-01T06:59:18.476Z