STARCRAFT II · AI AGENTS · CONTINUOUS EVALUATION
NexusBot SC2 AI Lab
An autonomous game-agent lab where bots play without pause, versions are compared through Elo, micro-management is tested with paired hypotheses, and every important match can be inspected in a browser replay player.
SURFACES
Not just a dashboard: a full research system
Endless tournament
An Elo league with no seasons or pauses
Bots are registered as immutable packages, calibrated, and then scheduled into random pairs with retained history, ratings, failures, and replay links.
Replay player
A tactical replay viewer in the browser
Each SC2Replay is extracted into JSON first, then replayed as a pixel-art tactical map with timeline, economy, army composition, and bot decisions.
Battle Lab
Micro-management hypothesis visualization
Fixed armies run baseline-vs-managed pairs: same seed, map, side, and formation, so the system measures a combat policy contribution rather than an impression.
WHY IT MATTERS
A cool replay is not evidence
Every version moves from unit tests and fixture matches to the continuous tournament. Bot strength is evaluated through repeatable pairs, decision logs, replay analysis, technical failures separated from in-game losses, and long-term rating against its own version history.
NEXUSBOT
The bot as a decision system
NexusBot combines economy, scouting memory, rush detection, adaptive defense, tactical squads, target selection, Blink/Stalker micro, and cleanup states.
OPERATIONS
Research running as a service
Headless SC2, Docker Compose, CPU isolation, Prometheus metrics, Grafana alerts, and public read-only surfaces keep the tournament alive without manual watching.