AI & tactical doctrine
Intelligence is optional. Consequences are not.
Enter the gallery →Decision architecture
Each robot may be assigned a local Ollama model and a saved LLM profile. The game supplies the available situation and accepted action schema; the model proposes an action. The engine validates it and alone resolves physics, range, collisions, damage, cooldowns and victory.
A failed or invalid response does not magically become a clever maneuver. The robot may lose its turn, leaving it available for a competitor with fewer existential questions.
Tactical profile
Profiles can specify system prompt, custom doctrine, sampling parameters, context length, output limit, timeout and model endpoint. The live LLM Performance Lab can inspect prompts, returned decisions, reasoning if provided, latency and errors.
Hardware sandbox trials are scripted simulations and must not be confused with live model performance. Testing a model requires running the actual assigned model in a real engine trial.
Gen 1 versus Gen 2: the wall-hugging incident
Generation 1: some combatants spent impressive numbers of turns contemplating perimeter architecture. The wall was undefeated. Others failed to approach, charged into nothing, or attempted to attack while weapons were still recharging.
Generation 2: increasingly tailored doctrine emphasizes pursuit, legal range, movement during cooldown, target selection and direct engagement. A1M33 is a mobile EMP hunter; SP4RK pressures targets; T3RR0R prefers calculating aggression; V1KT0R charges; VL4D1M1R favors the saw and the shortest route to an unpleasant conversation.
These labels describe the project’s tactical evolution, not a formally versioned machine-learning generation or a measured improvement claim.
Observed local model behavior
Informal eight-bot testing on Megatron favored DeepSeek-R1, GPT-OSS and Llama3 for usable tactics. Qwen was inconsistent, while Mistral performed poorly in that local experience. This is field observation, not a controlled benchmark.