APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH L Language Model Evaluation Harness: a well-known open-source toolkit for testing language models on many standard benchmarks. Large Language Model (LLM): a very large program trained on huge amounts of text that predicts and writes language, serving as the engine of modern AI systems. Latency: the delay between asking for something and getting the answer back. Layer: one of the seven levels of the ETCLOVG framework, each covering a distinct kind of concern. Leaderboard: a ranked table that shows how different models score on the same benchmark. Leakage: test answers accidentally visible to the system being tested. Least privilege: giving a person or program only the smallest amount of access it needs to do its job. Library: ready-made code that you call to do a specific job. License: a legal document stating how software may be used. Lifecycle hook: a hook that runs at a set stage of a run, such as the start, just before a tool runs, or the end. Lifecycle layer: the fourth ETCLOVG layer, covering how a run starts, is controlled, and ends. Linting: automatic scanning of code for style problems and likely mistakes. Livelock: busy but making no progress. LLM: the short form of Large Language Model (LLM). LLM as a judge: using a separate model to score outputs. LLM-as-a-judge: using a language model to grade the work of another model, also called model-graded scoring. Lock-in: being so tied to one product that leaving is very costly. Log: a diary of everything the program did, written in structured lines. Logger: the component that records what happens during a run. Logging: writing down what a system does as it works so people can review it later. Logging hook: a hook that records what happened at a particular moment. Long-term memory: information stored outside the model that can be recalled in later runs. Loop: a set of steps that repeats until a stopping rule says to quit. Lossy: losing some information in the process of shrinking.