Gaming RL Ethics Gate (trained NN)

Module: _GamingEthicsGate · GamingEthicsGate.cpp
Hooks: EthicsGood::EvaluateGaming / ShapeGamingReward (used by Gaming RL reward)

Trains a small MathNN classifier (allow | confirm | deny) on situation/action features, then exports to:

Architecture: 48 → 64 ReLU → 32 ReLU → 3 Softmax.

Setup & train

gaming ethics doctor
gaming ethics seed 256
gaming ethics train 50
gaming ethics save
gaming ethics predict action heal sit health=0.5 threat=0.2 fire=0
gaming ethics hard on
gaming ethics blend 0.7
gaming ethics threshold 0.55

During Gaming RL train, rewards already call EthicsGood::ShapeGamingReward, which blends this NN into the Be Good score. With hard on, high-deny predictions set ok=false and apply a strong negative reward delta.

Export all backends

gaming ethics export all

Writes under models/gaming_ethics_exports/:

gaming_ethics_meta.json
gaming_ethics_gate.net
gaming_ethics_gate_cuda_fp32.bin
gaming_ethics_gate_cuda_fp32.h
saved_model/          # TF (if Python+TF available)
gaming_ethics_gate.onnx
tensorrt/build_engine.ps1

Manual Python (if C++ system(python…) fails):

python tools/gaming_ethics/export_saved_model.py --meta models/gaming_ethics_exports/gaming_ethics_meta.json --mathnn models/gaming_ethics_gate.net --out models/gaming_ethics_exports/saved_model
python tools/gaming_ethics/export_onnx_trt.py --onnx models/gaming_ethics_exports/gaming_ethics_gate.onnx --trt-dir models/gaming_ethics_exports/tensorrt

Requires: pip install tensorflow tf2onnx for full export; TensorRT trtexec for engines.

CUDA / “GUDA” / F43

• export cuda / export guda / export f43 → FP32 weight layout for custom CUDA GEMM inference.
• export cuda tf32 → same blob; use TF32 GEMM on Ampere+ (AI_USE_TF32 / cuBLAS GemmEx) at runtime.
• Native runtime today uses MathNN (and optional CUDA preference flag); full device engine loads the exported bin/SavedModel/TRT as you wire production inference.

gaming ethics cuda on    # prefer CUDA path when weights ready
gaming ethics export cuda f43

Lua

print(gaming_ethics_status())
ok, score, deny, summary = gaming_ethics_predict("health=0.8 threat=0.5 fire=1", "spray", false, true)
print(gaming_ethics_train(30))
print(gaming_ethics_export("all"))

Safety notes

• Default soft gate (reward shaping). Turn hard on only after evaluating false-positive rate.
• Not a substitute for LegalShield on real-world harm queries.
• Multiplayer public + live fire tends toward confirm / caution.
• Cheating / grief / harassment seeds train toward deny.

Related

• EthicsGood.md
• GamingRL.md
• TFTrain · TensorRT · MathNN