SYSTEM ARCHITECTURE

Game AI as a production system

The project works as a continuous development loop: bot logic, immutable packages, headless SC2, rating, replay extraction, hypothesis lab, metrics, and public read-only interfaces.

Mobile view of the live NexusBot tournament

LAYERS

What the lab is made of

01

NexusBot core

The Protoss agent makes decisions about economy, scouting, defense, army grouping, target selection, and combat micro-control.

02

Immutable bot packages

Each version is registered as a separate package with a description, category, creation date, and difference from the previous generation.

03

Continuous tournament

The runner schedules pairs, launches headless SC2, records outcomes, technical failures, replay files, and updated Elo ratings.

04

Replay extraction

After each game, `.SC2Replay` becomes JSON with tracker events, positions, resources, units, buildings, and timeline events.

05

Battle Lab

A separate bench runs controlled army fights to test micro-management hypotheses before they reach the main bot.

06

Observability

Prometheus metrics, health endpoints, failure accounting, and public panels expose system state without server access.

OPERATING PRINCIPLE

Every new idea must leave a trail

A change is not considered successful until it has artifacts: a test, tournament rating signal, replay review, Battle Lab metrics, or an explainable failure. That keeps the project evolving as research infrastructure, not as one large game script.

PUBLIC READ-ONLY

The site is safe to browse

The public layer shows ratings, matches, the replay viewer, and lab results, but exposes no runner controls, secrets, or server-side tasks.

SCALING PATH

More games, more signal

The system is designed to expand the bot pool, scenarios, and maps: the more repeatable matches it runs, the clearer it becomes which decisions actually strengthen the agent.