APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH N Narrow AI: AI built to do a limited set of tasks well, which describes every real system in use today. Needle-in-a-haystack test: a test that buries one fact in a long text and checks whether the model can retrieve it. Needle-in-a-haystack testing: hiding one fact inside a very long text and checking whether the model can find it. Network egress control: a limit on which outside places an agent's program is allowed to send data to or reach. Neural network: a structure of many linked numerical units whose link strengths are adjusted during learning. Node: one dot in a graph, standing for one step or decision in a process. Normal case: an everyday typical request. Normal, edge, and adversarial cases: the three kinds of eval item. Normalization: cleaning text into a standard form before comparing.