APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH G Gate: an automatic checkpoint a change must pass. Gateway: a middle station that all model calls pass through, where they can be routed, saved for reuse, and counted. General AI: a so-called human-level kind of intelligence able to handle any task, which no system today actually is. Goal object: a structured description of the goal, constraints, inputs, and expected output. Golden answer: the trusted correct answer written in advance, used as the standard to grade a system against. Golden reference answer: a pre-verified correct answer stored beside a test input. Governance: the rules and controls that keep a system safe, secure, and lawful. Governance layer: the seventh ETCLOVG layer, covering safety, permissions, and rules compliance. Governance team: the team that owns rules about permissions, data, and approvals. GPU: the short form of Graphics Processing Unit (GPU). Graceful degradation: a system's ability to offer less than full service while still running, instead of failing completely. Graceful failure: ending a single run cleanly with a clear report when something goes wrong. Graceful termination: stopping in a planned, tidy way. Graph: a way of drawing logic as dots called nodes joined by arrows called edges. Graph, node, edge: points connected by lines that define allowed paths. Graphics Processing Unit (GPU): a chip built to do huge amounts of simple math at once, which makes it well suited to running language models. Grounding: tying a model's answer to real source material so it is less likely to make things up. Guard: a harness rule that stops or redirects bad behavior. Guardrail: a rule or check that keeps an agent's behavior inside safe limits.