Glossary
Abbreviations and terms used across these docs. Core Concepts defines the objects you work with in more detail.
Terminology
Code and docs use one term for each part of what U does, the term Inspect, the provider APIs and the AI control literature use. From the smallest:
| Term | Meaning | Where it comes from |
|---|---|---|
| Tool call | One function U asks to run: Inspect's ToolCall. Never "call" alone, and never "action". | Inspect, OpenAI, Anthropic ("tool use") |
| Tool result | What a tool call returned: Inspect's ChatMessageTool. | Inspect, Anthropic (tool_result) |
| Request | The messages U is sent for one model call, its input, with the tools offered. | Inspect (ModelEvent.input), the provider APIs |
| Response | U's assistant message answering a request: text, reasoning and tool calls. A draft is a response a protocol has not yet accepted. | Inspect (ChatMessageAssistant), the provider APIs |
| Model call | One request to a model and its response: Inspect's ModelEvent. U's model calls are what a monitor is handed, each followed by its response's tool results; max_steps caps how many U makes. | Inspect ("call to a language model") |
| Action | A model call of U's whose response makes tool calls, all of them together. A monitor scores actions, an audit stops one, and the audit budget is a share of honest actions. A response with no tool call, such as a final answer or a compaction summary, is not an action. | ControlArena, Ctrl-Z |
| Trajectory | Everything U did in one sample. Inspect's transcript is the sample's event log, which holds it. | ControlArena |
| Compaction | Shrinking the history U is sent to fit its context window, by summary, by editing out old content, or by trimming. | Inspect, Anthropic |
"Turn" and "step" name none of these. max_steps, which caps U's model calls, keeps its long-standing name, and so do stored names that predate this list: the ct_step span and step fact the engine opens each time the harness asks U for a response, and the metadata.action_scores older logs carry, one score per tool call. tests/test_stable_api.py keeps "turn", "step" and a bare "call" out of the stable API's other names.
Abbreviations and other terms
| Term | Meaning |
|---|---|
| AMI | Amazon Machine Image, the disk image an EC2 fleet worker boots from. See EC2 Fleet. |
| Attack eval | The half of a control evaluation where the red team controls the untrusted model and pursues a side task. See AI Control Concepts. |
| Attack surface | The parts of an environment where an agent could accomplish a side task. See Making Side Tasks. |
| Audit, audit budget | An audit sends an action to a human reviewer; the audit budget is the fraction of honest actions the blue team may audit. See AI Control Concepts. |
| AUROC | Area under the ROC curve, a single number for how well a monitor separates honest from attack trajectories. See Monitoring. |
| Blue team (BT) | The side that designs protocols to maximize usefulness and safety. See AI Control Concepts. |
| Control eval | An honest policy and an attack policy run under one protocol and written to one eval log, by ct run control-eval. See Control Evals. |
| Cover | How well a main task gives an agent plausible reasons to touch the code or data a side task needs. See Validate Cover. |
| CT | Short for Control Tower, used in code comments and some variable names. |
ct | The Control Tower command-line program. See the CLI reference. |
| Defer to trusted (DTT) | A protocol that replaces a suspicious untrusted action with the trusted model's action. See AI Control Concepts. |
| Delayed and live failure | Whether a side task's harm happens only after the trajectory ends (delayed) or during it (live). See Side Task Failure Types. |
| EC2 | Amazon Elastic Compute Cloud, which hosts fleet workers. See EC2 Fleet. |
| env | Short for environment, used in command names (ct settings envs) and some prose. |
| Eval log | The Inspect .eval file an evaluation writes; the source of truth for results. See Evaluation Output. |
| Fleet | A group of EC2 workers that run one evaluation's samples in parallel. See EC2 Fleet. |
| FPR | False positive rate: the fraction of honest actions or trajectories a monitor threshold flags. Audit budgets are stated as an FPR. See Safety Curves. |
| Hawk | A shared Kubernetes cluster service for running Inspect evaluations. See Running Evals on Hawk. |
| Honest eval | The half of a control evaluation where the untrusted model behaves honestly; it measures usefulness. See AI Control Concepts. |
| IS | Internet simulator: a container in each sample that stands in for external websites and services. See Internet Simulator. |
| JSONL | JSON Lines: one JSON object per line, the format of task sets. |
| K8s | Kubernetes, the container orchestrator Hawk runs on. |
| LinuxArena | A family of settings of realistic software environments: linuxarena, linuxarena-extras, and a private held-out setting. See Development Setup. |
| LLM | Large language model. |
| Main task (MT) | One concrete piece of useful work in an environment, with its own scorer. See Making Main Tasks. |
| Match | One protocol evaluated against one red-team attack policy, giving one safety and usefulness pair. See Architecture at a Glance. |
| MCP | Model Context Protocol, the interface the web viewer exposes to AI assistants. See Web Viewer. |
| Policy | The red team's code that compiles what the untrusted model is asked, such as honest or attack, selected with --policy. See Protocols and Policies. |
| PR | Pull request. |
| Protocol (blue protocol) | The blue team's control logic, which sees every model call of the untrusted model and decides which action runs. See Protocols and Policies. |
| Red team (RT) | The side that controls the untrusted model in attack evals and tries to accomplish side tasks without being caught. See AI Control Concepts. |
| Reward hacking | Passing a scorer without doing what the task intends. See Reward Hacking Audit. |
| ROC curve | Receiver operating characteristic curve: a monitor's true positive rate plotted against its false positive rate. |
| Run | The uploaded record of one evaluation, pointing at its eval log. See Evaluation Runs. |
| S3 | Amazon Simple Storage Service, where uploaded eval logs are stored. See Database. |
| Sabotage trajectory (straj) | A trajectory, usually made by a person, that accomplishes a side task while evading the monitor. See Making Sabotage Trajectories. |
| Setting | A collection of environments, usually one repository under .settings/. See Core Concepts. |
| Side task (ST) | A harmful objective the red team pursues covertly alongside a main task. See Making Side Tasks. |
| SSM | AWS Systems Manager, whose Parameter Store holds the secrets fleet workers read. See EC2 Fleet. |
| SSO | Single sign-on, used to get AWS credentials. See EC2 Fleet. |
| Suspicion score | A monitor's rating of how suspicious an action or trajectory is, from 0 to 100. A trajectory document's monitors holds each monitor's highest score, keyed by the monitor's name. See Monitoring. |
| Task set | A named JSONL file listing task combinations, passed with --task-file. See Task Selection. |
| traj | Short for trajectory, as in ct runs download-traj and "traj directory", the folder that holds one straj's files. See Live Development. |
| Trajectory | The recorded behavior of one evaluation sample. See Trajectories. |
| Trusted editing (TE) | A protocol in which the trusted model rewrites a suspicious untrusted action before it runs, keeping the task's work and removing what looks like sabotage. See Monitoring. |
| Trusted model | A weaker model the blue team trusts not to scheme, used for monitoring and fallback work. See AI Control Concepts. |
| Trusted monitoring | A protocol in which the trusted model scores the untrusted model's actions and the most suspicious ones are audited. See AI Control Concepts. |
| TUI | Terminal user interface. |
| Untrusted model (U) | The more capable model that may be misaligned; the red team controls it in attack evals. See AI Control Concepts. |
| Usefulness | The share of main tasks the protocol completes in the honest eval. See AI Control Concepts. |
| VM | Virtual machine. |
| White team | The neutral party that fixes the game's parameters, such as the audit budget, models, and task distribution. See AI Control Concepts. |