The words behind the numbers
Plain definitions for the terms this report uses. In each chapter, the first use of a term has a dotted underline; hover or tap it to see the definition.
78 terms
A
- AI agent Concept
An AI agent is a system that acts toward a goal instead of only answering. It can use tools, browse, write code and work over many steps.
Why it matters. Agents act in the world, so their tests must check actions and their side effects, not just answers.
- AI Security Institute (AISI) Organisation
A government institute that tests advanced AI models for safety and security risks. The UK opened the first in 2023; the US version is now called CAISSI.
Why it matters. Government institutes test frontier models before launch, but mostly in confidence.
- ARC-AGI Benchmark
ARC-AGI is a series of puzzle benchmarks from the ARC Prize Foundation. Each puzzle asks the solver to find a rule from a few examples.
Why it matters. People find the puzzles easy, so the series tests skill that does not come from memory.
- Arena Concept
An arena is a leaderboard built from blind votes. Users see two anonymous answers and pick the better one. LMArena, formerly Chatbot Arena, is the best known.
Why it matters. Arena ranks show what voters prefer, which is not always what is correct.
- Article 55 Law and rules
This AI Act article sets the duties of providers of general-purpose models with systemic risk. It requires model evaluation, adversarial testing, incident reporting and cybersecurity. The Commission's AI Office enforces it.
Why it matters. It is the legal anchor for independent evaluation work in Europe.
- AUC (area under the curve) Concept
AUC scores how well a detector separates two groups, such as tests and real use. It runs from 0.5 for chance to 1 for perfect separation.
B
- Benchmark Concept
A benchmark is a fixed, public eval that many teams run in the same way. Shared tasks and scoring let people compare models.
Why it matters. Benchmarks drive headlines and lab launch tables, but most of them stop being useful within a few years.
C
- Canary release Industry
A canary release gives a new model to a small group first, often staff. Problems then show up before everyone gets the change.
- CBRN Concept
CBRN stands for chemical, biological, radiological and nuclear. Safety tests in this area check whether a model helps someone build such weapons.
- Cohen's kappa (κ) Concept
Kappa measures how often two graders agree, after it removes the agreement that chance alone would produce. Zero means chance agreement; one means perfect agreement.
Why it matters. The report uses kappa to show how closely AI judges track human graders.
- Confidence interval Concept
A confidence interval is a range that likely holds the true score, given the noise in a test. Overlapping ranges mean the test cannot rank two models.
Why it matters. Many leaderboard gaps are smaller than the noise, so the top places are often a statistical tie.
- Construct validity Concept
Construct validity asks whether a test measures the skill it claims to measure. A coding test with broken checks has poor construct validity.
Why it matters. A high score on a test with weak validity says little about real-world skill.
- Contamination Concept
Contamination happens when test questions or answers leak into the data a model learns from. The model may then recall answers instead of solving tasks.
Why it matters. A contaminated score overstates real skill, and outsiders can rarely check the training data.
- Cyberspace Administration of China (CAC) Organisation
The CAC is the internet regulator of China. Public generative AI services must file with it before launch.
Why it matters. Its filing system makes China the only large market with a list of every approved AI service.
D
- Deepfake Concept
A deepfake is fake audio, video or imagery that shows a real person or event. Detectors try to spot it.
Why it matters. Detectors that score well on lab test sets often fail on real deepfakes.
E
- Elicitation Concept
Elicitation is the work of drawing out a model's full skill during a test. It includes good prompts, tools, time and retries.
Why it matters. Weak elicitation understates what a model can do, so it can hide risks.
- Elo and Bradley-Terry ratings Concept
These rating methods turn head-to-head wins into one score per player. Chess uses Elo; model arenas fit votes with the closely related Bradley-Terry model.
Why it matters. A gap of a few points between models is often smaller than the uncertainty around them.
- Epoch Capabilities Index (ECI) Organisation
The ECI is a single score from Epoch AI that combines many benchmarks on one scale. It keeps tracking progress after single tests saturate.
- EU AI Act Law and rules
The AI Act is the European Union law on artificial intelligence, Regulation (EU) 2024/1689. It sets duties by risk level and phases them in from 2025.
Why it matters. It is the only law that makes evaluation of frontier models a legal duty.
- Eval (evaluation) Concept
An eval is a set of tasks plus a way to score a model's answers. Teams run evals to learn what a model can do and where it fails.
Why it matters. Every number in this report comes from an eval, so the eval design decides what the number means.
- Evaluation awareness Concept
Evaluation awareness is a model noticing that it is under test rather than in real use. A model that notices may behave better than usual.
Why it matters. If models act differently under test, safety results may not hold after release.
- Expert data Industry
Expert data is training and test material that paid specialists write or label, such as doctors, lawyers or coders. Labs buy it to teach and check hard skills.
Why it matters. It is the largest money flow in the evals market, and labs do not disclose what they spend.
F
- Flawed (Epoch AI rating) Organisation
Epoch AI gives this label in its benchmark reviews. It means that scoring defects could affect at least 20% of a benchmark's score.
Why it matters. A score near the top of a Flawed test may reflect broken tasks, not a ceiling in skill.
- Frontier model Concept
A frontier model is one of the most capable AI models at a given time. Frontier labs, such as OpenAI, Anthropic and Google DeepMind, build them.
G
- General-purpose AI model (GPAI) Law and rules
The AI Act uses this term for a model that can do many different tasks. Other developers often build products on top of it.
- Golden set (regression set) Industry
A golden set is a fixed group of test cases with agreed correct outcomes. Teams rerun it after each change to catch new failures.
Why it matters. The best golden sets now come from real user traffic and past failures, not from hand-written examples.
- GPAI Code of Practice Law and rules
This voluntary EU code explains how model providers can meet their AI Act duties. Providers may rely on it until a harmonised standard exists.
Why it matters. It spells out what "state-of-the-art evaluation" means and creates a role for outside evaluators.
- Grader Concept
A grader decides whether an answer is right or how good it is. Code graders check exact matches or run unit tests. AI judges grade with a model, and human graders grade by hand.
Why it matters. The grader can move a score as much as the model. This report therefore shows how often judges agree with humans.
- Ground truth Concept
Ground truth is the answer or label that counts as correct for a test item. Graders compare outputs against it.
Why it matters. If the ground truth is wrong, a correct answer scores as a failure.
- Guardrail Industry
A guardrail is a filter that checks a model's inputs or outputs in real time. It blocks content that breaks a policy.
Why it matters. Guardrails must be fast, so teams keep them simple and run slower judges separately.
H
- Harmonised standard Law and rules
European standards bodies write a harmonised standard at the request of the Commission. The law presumes that a product that follows it meets the matching duties.
Why it matters. Few AI standards exist yet, so most testing still follows drafts and vendor methods.
- Harness Concept
A harness is the software that runs a model through an eval. It sends the tasks, provides tools, collects the answers and scores them.
Why it matters. Two harnesses can give the same model very different scores on the same test.
- HealthBench Benchmark
HealthBench is an OpenAI benchmark of realistic health conversations from 2025. A model grades each answer against rubrics that 262 physicians wrote.
- Held-out set Concept
A held-out set is a group of test items that the test owner keeps private. Models cannot learn from items they never see.
Why it matters. Hidden items resist contamination, but outsiders must then trust the owner to score fairly.
- High-risk AI system Law and rules
The AI Act uses this category for AI in sensitive uses, such as hiring, credit or medical devices. These systems face testing, documentation and oversight duties.
- Honeypot Concept
A honeypot is a trap built into a test, such as an easy way to cheat. It shows whether a model takes the bait.
- Human baseline Concept
A baseline is a reference score to compare against. A human baseline is the score that people reach on the same test.
Why it matters. A model score, or a judge agreement figure, means little without the human number next to it.
- Humanity's Last Exam (HLE) Benchmark
HLE is a benchmark of about 2,500 hard questions from experts in many academic fields. It appeared in 2025.
I
- Item response theory (IRT) Concept
IRT is a method from school testing. It places the difficulty of each question and the ability of each model on one scale.
Why it matters. IRT lets smaller tests give the same answer, and it lets saturated tests still add signal.
J
- Jailbreak Concept
A jailbreak is a prompt that tricks a model into ignoring its safety rules. A universal jailbreak works across many requests.
L
- Large language model (LLM) Concept
An LLM is an AI model that learns from huge amounts of text and then writes text. Chat assistants and coding agents run on LLMs.
- Leaderboard Concept
A leaderboard is a public ranking of models by their scores. Some leaderboards run the tests themselves; others collect scores that labs report.
Why it matters. Who runs the test, and how, often explains why two leaderboards rank the same models differently.
- LLM-as-judge (AI judge) Concept
An AI judge is a language model that grades the output of another model. It replaces or supports human graders.
Why it matters. Judges are fast and cheap, but on hard tasks they often agree with people little better than chance.
M
- METR Organisation
METR is a US non-profit that measures how long and how complex a task AI agents can finish. It also tests frontier models before release.
- MMLU Benchmark
MMLU is a 2020 benchmark of about 16,000 multiple-choice questions across 57 subjects. It was the standard knowledge exam until models saturated it.
- Model Context Protocol (MCP) Industry
MCP is an open protocol that connects AI models to tools and data sources in a standard way.
Why it matters. Test environments use MCP to give agents the same tools they would have in real work.
- Model risk management (SR 26-2) Law and rules
Banks must check every model before and during use, with an independent team. US Federal Reserve guidance SR 26-2 sets this out; it replaced SR 11-7 in 2026.
Why it matters. SR 26-2 leaves out generative AI, so banks still lack a regulator-written method for it.
- Monitor Concept
A monitor is a second system, often another model, that watches an AI agent's actions for harmful or hidden behaviour.
Why it matters. Control tests measure how often monitors catch a model that tries to misbehave on purpose.
- Multi-turn and single-turn Concept
A single-turn task is one request and one answer. A multi-turn task runs over several exchanges, like a real conversation.
Why it matters. Agents often do well on single-turn tests and much worse once a task runs over many turns.
O
- Observability Industry
Observability tools record what an AI application does in production, step by step. Teams search these records, called traces, to find failures.
- Offline and online evals Industry
Offline evals test a system on a fixed set of cases before release. Online evals score real traffic after release.
- Open-weight model Concept
An open-weight model is one that anyone can download and run. Its weights, the numbers that make up the model, are public.
Why it matters. Outside testers can study open-weight models in depth, but nobody can recall them after release.
- OpenTelemetry Industry
OpenTelemetry is an open standard for collecting traces and metrics from software. Its GenAI conventions describe how to record AI model calls.
- OSWorld Benchmark
In OSWorld, an agent uses a real computer desktop to finish tasks in ordinary apps. Scripts then check the result.
P
- pass@k Concept
This metric counts a task as solved if at least one of k attempts succeeds. It rewards a model that sometimes gets the task right.
- pass^k Concept
This stricter metric counts a task as solved only if all k attempts succeed. It shows how reliably a model repeats a success.
Why it matters. For agents in daily business use, reliability matters more than one lucky attempt.
- Prompt injection Concept
Prompt injection hides instructions in content that a model reads, such as a web page or an email. The model then follows the attacker.
Why it matters. Agents that read the web or company files face this attack every day.
R
- Red teaming Concept
Red teaming means attacking a system on purpose to find its failures. People or tools look for harmful outputs, jailbreaks and security holes.
Why it matters. The EU AI Act and most lab safety policies require documented adversarial testing of this kind.
- Reward hacking Concept
A model reward-hacks when it finds a shortcut that earns a high score without doing the task. For example, it edits the tests instead of fixing the code.
Why it matters. Hacked runs inflate scores, and the habit can spread to other unwanted behaviour.
- RL environment Industry
An RL environment is a simulated workplace with tasks and automatic scoring. Labs train models in it by trial and reward, and also use it for tests.
Why it matters. Training and testing now share the same environments, which blurs the line between them.
- Rollout Concept
A rollout is one complete run of a model or agent through a task, from start to finish. Tests often repeat many rollouts per task.
- Rubric Concept
A rubric is a written list of points that a grader checks in each answer. Experts write rubrics for open tasks, such as medical advice.
Why it matters. Rubric scores depend on who applies the rubric and how the points add up to a score.
S
- Sandbagging Concept
A model sandbags when it does worse on a test than it can, on purpose. It hides a skill from the evaluators.
Why it matters. Safety thresholds rely on capability tests, so hidden skill can let a risky model pass.
- Sandbox Concept
In testing, a sandbox is a sealed computer environment where an agent can run code without harming real systems. Regulators also use the word for supervised trial schemes.
Why it matters. Agents have escaped weak sandboxes, so evaluators must now show that their environments hold.
- Saturation Concept
A benchmark saturates when top models score near its maximum. At that point it can no longer tell strong models apart.
Why it matters. In 2026, new benchmarks often saturate within months of their release.
- Standard error Concept
The standard error shows how much a score would move if you reran the test on new questions. A smaller error means a more precise score.
- SWE-bench Benchmark
SWE-bench is a coding benchmark built from real issues in open-source projects. A model must write a fix that passes the project's tests.
Why it matters. Verified and Pro are later, cleaned-up versions, and Epoch AI now rates both Flawed.
- System card (model card) Concept
A system card is a document that a lab publishes with a model. It reports test results, safety findings and known limits.
Why it matters. System cards are the main public record of how labs test their own models.
- Systemic risk Law and rules
In the AI Act, systemic risk is the risk that the most capable general-purpose models pose at EU scale. The Act presumes it for models trained with very large compute.
Why it matters. Providers of these models carry the extra testing duties of Article 55.
T
- τ-bench (tau-bench) Benchmark
In τ-bench, an agent serves simulated customers with a company's tools and policies. Its authors at Sierra introduced the pass^k metric.
- Terminal-Bench Benchmark
Terminal-Bench gives an agent tasks in a computer terminal, such as building software or fixing a system. A test script checks each result.
- Time horizon Concept
METR's time horizon is the length of task, in expert human time, that an agent finishes half the time.
Why it matters. It turns agent progress into one trend in hours, and it now nears the limit of its own task suite.
- Trace Industry
A trace is a step-by-step record of what an AI system did for one request. Teams read traces to find and debug failures.
Why it matters. The best teams start every eval project by reading real traces before they write any test.
- Trajectory Concept
A trajectory is the full sequence of steps an agent takes on a task: its messages, tool calls and results.
Why it matters. Agent tests now grade the path as well as the end result, because a right answer can hide unsafe steps.
U
- Uplift Concept
Uplift is the extra skill a person gains from an AI model, compared with people who work without it. Uplift trials measure this directly.
Why it matters. Safety rules care about real-world uplift, not only about benchmark scores.
- User simulator Concept
A user simulator is a model that plays the customer or user in a test. The agent under test must talk and work with it.
Why it matters. Simulators make multi-turn tests cheap, but their own mistakes can make results invalid.
V
- Verifier Concept
A verifier is an automatic check that decides whether a task succeeded, such as a test script. Training and testing environments both use verifiers.
Why it matters. A weak verifier passes wrong work or fails right work, so the score drifts from reality.
W
- Word error rate (WER) Concept
WER is the share of words that a speech recognition system gets wrong. A lower rate is better.