Psycho uses a GRU neural network trained on real rotation data to detect aim assistance with <0.01 false-positive rate. No signatures. No bypasses. Pure pattern recognition.
Real metrics from production servers running Psycho.
A Gated Recurrent Unit trained on real player data. Learns temporal rotation patterns invisible to rule-based checks.
No hardcoded thresholds or delta limits. The model generalises to new cheat clients without updates.
Auto-computed class weights at train time. Imbalanced datasets train correctly out of the box.
Collect data on your server, run /psycho train, deploy. No Python required.
Pure Java forward pass with pre-allocated arrays. Zero GC pressure on the hot path. Scales to 300+ players.
VL thresholds, decay rates, punishment commands — everything in config. No recompilation needed.
Every check runs independently with its own VL buffer, threshold and punishment command.
All commands require the psycho.admin permission unless noted.
| Command | Description | Permission |
|---|---|---|
| /psycho train <epochs> <lr> | Train the GRU model on the current dataset. | psycho.admin |
| /psycho collect <player> <legit|cheat|stop> | Start or stop recording rotation data for a player. Use stop to end the session. |
psycho.admin |
| /psycho reload | Reload config files without restarting the server. | psycho.admin |
Requires Paper 1.14+ and PacketEvents 2.x as a dependency.