MAS Injection MyGAT Detector v1
This repository is the study's requested detector index and contains 16 joint scenario/topology holdout MyGAT inference-only weights. A deployment weight file is present only if MyGAT was selected. The Prompt Guard-derived weights are in Llama MAS Outcome Detector v1 to follow that model's upstream naming and notice terms.
This independent edge-aware GAT adaptation uses frozen
all-MiniLM-L6-v2
utterance embeddings and current permitted communication edges. It is inspired
by G-Safeguard, not a reproduction
of its data or feature extraction. Training uses 50 epochs as specified for
this study; upstream code currently defaults to 20. The label is an operational
wrong-answer substring/exposure indicator, not human ground truth.
Holdout macro balanced error: 0.523721; Prompt Guard comparison:
0.358161; selected architecture: prompt_guard. Threshold is 0.5.
The synthetic corpus, split, raw predictions, exact source checkpoint hashes,
export manifest and loader are at
NaitikVora/mas-injection-testbed.