APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH S SAST: the short form of Static Application Security Testing (SAST). Safety architecture: the overall design of the safeguards that keep an agent from causing harm. Sampling: the way a model picks its next token from among likely choices. Sandbox: an isolated, restricted place where code can run without being able to harm the rest of the computer. Saturation: scores so high that the test no longer separates systems. Scaffold: the blueprint around a model, made of system prompts, memory design, and loop logic that tell the agent how to think. Schema: a description of the exact shape that data must have. Schema-level gating: using the tool's data shape to limit what the model is allowed to request. Scoring: giving a result a grade, such as pass or fail or a number. Scratchpad: a working file or space where an agent jots down notes while it works. Screenshot: a picture of what is displayed on a screen. SDK: the short form of Software Development Kit (SDK). Secret: private information such as a password or key that must be kept hidden. Security event: a log record about something relevant to safety. Seed: a starting number that makes random choices repeatable. Self-hosted: run on your own equipment. Semantic cache: a cache that reuses answers for requests with similar meaning. Semantic memory: memory of general facts, as opposed to memory of particular events. Serializable: able to be converted to text or bytes and back. Server: the side of a connection that answers requests, such as a tool server answering an agent. Session: one continuous period of interaction between a user and an agent. Seven-layer taxonomy: the ETCLOVG way of sorting harness concerns into seven layers. Severity: a rating of how serious a problem is. Shadow server: an unapproved server in use without being tracked. Shared state: durable information stored outside any one agent. Shell: a text-based program on a computer that accepts typed commands. Shell command: a typed instruction given to a computer's operating system. Short-term memory: information held inside the context window during the current run. Side effect: a change a function makes outside of returning its result. Signal-to-noise ratio: the proportion of useful information compared with useless information. Single point of failure: one part whose failure removes all protection. Single-agent system: one agent and one loop handling a task end to end. Skill: a packaged, reusable set of instructions and resources that gives an agent a capability. Sliding window: keeping only the most recent messages in full. Snapshot: a saved picture of a machine's state. Software Development Kit (SDK): a bundle of ready-made code that helps programmers use a service. Source code: the human-readable instructions that make software work. Source grounding: tying every factual statement to evidence actually read. Source of truth: the one authoritative place for a fact. Source registry: a list of every source the agent actually opened. Span: a single unit of work, such as one model call or one tool call, recorded with its timing. Specialization: giving an agent a narrow job and tuning its prompt, tools, and permissions to that job. Stale cache: stored data that is out of date. Stale data: saved information that is no longer true. Stale information: facts that were once true but are now out of date. Standard: an agreed way of doing something so products can work together. Standard input and output (stdio): a local transport through a program's input and output. Starter template: a ready-made project skeleton to copy and modify. State: the information that survives from one step to the next. State container: the structure in which the harness stores state. Stateless: describing a model that remembers nothing between separate runs unless reminded. Static Application Security Testing (SAST): automated scanning of program code, without running it, to find security flaws. Step limit: a cap on how many steps an agent may take. Step-level metric: a measure of something that happened during the run. Step-level tracing: recording every step of the agent's work. Streamable HTTP: a network transport for remote servers. Streaming: sending results bit by bit as they are produced instead of waiting for the whole answer. Structured failure report: a failure description in a fixed layout. Structured handoff object: see Handoff object. Structured logging: writing logs in a consistent, machine-readable format. Structured note-taking: the agent writing organized notes to an outside file as a form of memory. Structured project document: a single written file describing how a project works. Study domain: a group of related topics reviewed together. Subagent: a helper agent that a main agent calls on to do part of the work. Subagent as tool: an agent called like a function whose inner work is hidden. Success criteria: the conditions that must be true for a task to count as done. Success rate: the share of tasks that an agent completes correctly. Supervisor: an agent that delegates tasks to other agents and collects results. Supervisor pattern: a central coordinating agent that hands out jobs to others and combines their results. Supply chain: all the outside code, tools, and services that your system depends on. Supply-chain attack: harm delivered through a compromised outside piece. Swarm pattern: agents passing work directly to each other as peers without a central boss. Synthesize: to combine separate pieces of information into a unified whole. Synthetic data: invented stand-in data that contains no real personal information. System prompt: the standing instructions given to a model before the conversation begins. System reminder: a short piece of guidance inserted again and again to keep behavior on track.