The bots have a deliberately human sensorium: an awareness cone that has to
fill before they see you, a finite turn rate, reaction latency, cover
discipline, magazines they actually reload. Sneaking is real — 2.3 to 3.9
seconds of silence directly behind a bot, measured, against under 0.2 for a
loud frontal approach. The founding constraint of the whole bot programme: they
never became better shots. Their accuracy is byte-identical to the code before
the work began; everything they gained is decision quality.
Above that layer sits a learned policy. Which objective to take, when to
rotate, when to go home — those decisions come from a neural network trained by
reinforcement learning in a 36-hour self-play campaign, and the champion
policies from that campaign hold the production seats today. The policy chooses
where a bot goes. It is never allowed to touch the trigger.
Bot count and skill are live dials in OPTIONS. The skill slider scales the
human limitations — reaction, turn rate, notice speed — never aim, so the floor
is a clumsy squad, not a blind one.
The no-aimbot rule, the
training campaign, and the numbers — on the engineering page. →