// Evals 101 · Glossary

Jargon decoder

Every eval term in plain words, with an analogy and a line you’ll hear at work. Eval terms come first; general AI words sit under Related AI terms.

Showing 68 of 68 terms

Eval terms (42)

The words you need to build, run and read evals.

Alignment

Making an AI system’s behaviour match what its designers and users actually intend.

Analogy
Making sure the new hire understood the job, not just the words of the brief.
You’ll hear it
“That’s an alignment problem, not a capability problem.”

Benchmark

A public, standard test set used to compare models on general skills.

Analogy
A national exam: good for comparing schools, not for hiring for your job.
You’ll hear it
“It tops the coding benchmark, but fails our internal suite.”

Binary rubric (checklist)

Grading with yes/no questions instead of a numeric score.

Analogy
Pass/fail checkboxes versus “rate this 1 to 10”.
You’ll hear it
“Swap the 1–10 score for five yes/no checks.”

Source: Zhang et al. (2026)

Cohen’s kappa

A score for how much two graders agree, after removing agreement expected by pure chance.

Analogy
Two referees who both always call “no foul” agree a lot, but that proves little.
You’ll hear it
“What’s the kappa between the AI judge and our human labels?”

Source: McHugh (2012)

Data contamination

A model saw a benchmark’s test questions during training, so its score looks better than it is.

Analogy
A student who memorised last year’s exam paper.
You’ll hear it
“Those benchmark numbers might be contaminated.”

Source: Sainz et al. (2023)

Environment

A safe copy of real tools and data where an agent can practise and be tested.

Analogy
A flight simulator.
You’ll hear it
“Spin up a fresh environment for each run.”

Environment leakage

The expected answer is accidentally visible to the agent inside the test environment.

Analogy
A test paper handed out with the answers on the back.
You’ll hear it
“Scores jumped overnight — check for leakage first.”

Error analysis

Reading real failures one by one and sorting them into types before fixing anything.

Analogy
A doctor diagnosing before prescribing.
You’ll hear it
“Do error analysis on last week’s failed traces first.”

Eval

A repeatable check of whether an AI system did its job correctly.

Analogy
A driving test for the AI.
You’ll hear it
“Do we have evals for this feature before we ship it?”

Eval suite

A tagged collection of tasks, each with an instruction, an environment and a verifier.

Analogy
A full exam paper, split into subjects.
You’ll hear it
“The billing suite went red after the model swap.”

Evals flywheel

Production failures become new evals, which drive fixes, which you test and ship again.

Analogy
A gym routine that adapts to your last workout.
You’ll hear it
“Every support ticket feeds the flywheel.”

Fine-tuning

Training an existing model further on your own examples to specialise it.

Analogy
Sending a specialist on a short course.
You’ll hear it
“Fine-tune only after prompts and tools stop helping.”

Golden dataset

A small, hand-checked set of test cases the team trusts as the reference.

Analogy
The teacher’s own marked copy of the exam.
You’ll hear it
“Run it against the golden set before we compare models.”

Grounding

Tying a model’s answer to specific sources or data it was given, so claims can be checked.

Analogy
An essay with footnotes.
You’ll hear it
“Is that answer grounded in the policy document?”

Guardrails

Checks and rules around a model that block or fix unsafe or off-policy inputs and outputs.

Analogy
The barriers on a mountain road.
You’ll hear it
“The guardrail blocked the agent from emailing an external address.”

Harness engineering

Improving the prompts, tools and model choice wrapped around an agent. Not the same as an eval harness.

Analogy
Tuning the workplace, not the worker.
You’ll hear it
“Most of our gains came from harness engineering, not a new model.”

Human-in-the-loop

A person reviews, approves or corrects the AI at key points.

Analogy
An editor who signs off every article before it’s printed.
You’ll hear it
“Refunds over a set amount stay human-in-the-loop.”

Non-determinism

The same input can produce different outputs on different runs.

Analogy
Asking the same person the same question on different days.
You’ll hear it
“It passed yesterday and failed today with no code change.”

Observability

Being able to see what a live system is doing and why, from its logs, traces and metrics.

Analogy
The dashboard and dashcam in a car.
You’ll hear it
“We have no observability on the agent in production.”

OpenTelemetry

An open standard for collecting traces and metrics; its conventions for AI calls are still being written.

Analogy
A shared plug shape, so any device fits any socket.
You’ll hear it
“Emit OpenTelemetry spans so we can switch vendors later.”

Source: OpenTelemetry (2026)

Pairwise comparison

Grading by showing a judge two answers side by side and asking which is better.

Analogy
An eye test: “clearer with lens one, or lens two?”
You’ll hear it
“Run a pairwise comparison of the old prompt against the new one.”

Source: Zheng et al. (2023)

Pareto frontier

The options where you can’t get better quality without paying more, or pay less without losing quality.

Analogy
The best-value seats on a plane: nothing else is both cheaper and comfier.
You’ll hear it
“Model B sits on the frontier for internal tasks.”

Source: Kapoor et al. (2024)

pass@k

The chance that at least one of k attempts succeeds.

Analogy
A basketball player gets k shots; one basket counts as a pass.
You’ll hear it
“pass@3 looks great, but users only get one try.”

Source: Chen et al. (2021)

pass^k

The chance that all k attempts succeed. A measure of reliability.

Analogy
The player must sink every one of k free throws.
You’ll hear it
“For a customer-facing agent, watch pass^k, not pass@k.”

Source: Yao et al. (2024)

Precision

Of everything the AI flagged, the share that was truly right.

Analogy
A fishing net: how much of the catch is the fish you wanted, not old boots.
You’ll hear it
“Precision is high: when it flags a ticket as urgent, it usually is.”
Related
RecallEval

Recall

Of everything the AI should have found, the share it actually found.

Analogy
The same net: how many of the fish in the lake did you catch?
You’ll hear it
“Recall is low: it misses a lot of the urgent tickets.”

Red teaming

Deliberately attacking your own AI system to find ways it can be misused or broken.

Analogy
Hiring a locksmith to try to break into your house.
You’ll hear it
“Red team it for prompt injection before launch.”

Regression eval

Re-running existing evals after any change to catch things that used to work and now don’t.

Analogy
A smoke alarm in the build pipeline.
You’ll hear it
“The prompt tweak fixed tone but broke three regression tasks.”

Sandbox

An isolated, disposable space where code or agents can run without touching real systems.

Analogy
A terrarium: everything inside, nothing leaks out.
You’ll hear it
“Run the agent’s code in a sandbox, never on the host.”

Span

One timed step inside a trace, such as a single model call or tool call.

Analogy
One receipt in the stack.
You’ll hear it
“The search tool span is where the time goes.”

Synthetic data

Test or training examples generated by software or AI instead of collected from real use.

Analogy
Practice questions written by a tutor rather than taken from real exams.
You’ll hear it
“Synthetic cases are a start, but add real ones from production.”

Trajectory eval

Grading the path an agent took — its steps and tool calls — not just the final answer.

Analogy
Marking the working, not only the answer, in a maths exam.
You’ll hear it
“Right answer, but it deleted a record on the way. Trajectory eval caught it.”

Related AI terms (26)

General AI words you’ll hear around evals. Useful context, not required to start.

Agentic coding

AI agents that read, write, run and test code in a project, with a developer reviewing.

Analogy
A junior developer who works fast and needs code review.
You’ll hear it
“The agent opened a PR; CI and review still gate the merge.”

Agentic workflow

A process where an AI plans and carries out several steps, often with tools, toward a goal.

Analogy
A recipe the cook follows and adjusts as they go.
You’ll hear it
“We turned the onboarding checklist into an agentic workflow.”

AI-native SDLC

A loose label for software processes designed around AI agents doing parts of planning, coding and testing.

Analogy
Redesigning the kitchen around a new appliance, not just plugging it in.
You’ll hear it
“In an AI-native SDLC, evals are part of CI.”

Chain-of-thought

Getting a model to write out intermediate reasoning steps before the final answer.

Analogy
Thinking out loud.
You’ll hear it
“Ask for chain-of-thought, then grade only the final answer.”

Context engineering

Choosing what information, tools and history go into the model’s context for each step.

Analogy
Packing exactly the right papers for a meeting — not the whole filing cabinet.
You’ll hear it
“The prompt is fine; it’s a context engineering problem.”

Embeddings

Lists of numbers that represent meaning, so similar texts end up close together.

Analogy
Map coordinates for ideas: similar ideas live in the same neighbourhood.
You’ll hear it
“We search by embeddings, not keywords.”

Inference

Running a trained model to get an output. What happens every time you send a prompt.

Analogy
The chef cooking your order, as opposed to learning to cook.
You’ll hear it
“Inference cost is our biggest bill.”

Latency

How long you wait for a response.

Analogy
The time between ordering and the food arriving.
You’ll hear it
“The big model is better, but latency is too high for chat.”

LLM (large language model)

An AI model trained on huge amounts of text to predict and generate language.

Analogy
A very well-read autocomplete.
You’ll hear it
“Which LLM is behind this feature?”

LoRA

A cheaper fine-tuning method that trains small add-on pieces instead of the whole model.

Analogy
Adding a lens to a camera instead of building a new camera.
You’ll hear it
“We shipped a LoRA adapter for the support tone.”

Source: Hu et al. (2021)

MCP (Model Context Protocol)

An open protocol for connecting AI applications to tools and data sources in a standard way.

Analogy
A universal adapter plug for AI tools.
You’ll hear it
“Is there an MCP server for our CRM?”

Multi-agent system

Several AI agents, each with its own role, working together or handing work to each other.

Analogy
A kitchen brigade: one chops, one cooks, one plates.
You’ll hear it
“The writer agent hands drafts to the critic agent.”

Open-weight vs. closed model

Open-weight models publish their trained weights for download; closed models are only reachable through a provider.

Analogy
Buying a car versus using a taxi service.
You’ll hear it
“We need an open-weight model to run it on our own servers.”

Prompt engineering

Writing and refining prompts so a model does the job reliably.

Analogy
Rewriting a job brief until the freelancer gets it right first time.
You’ll hear it
“Try prompt engineering before you fine-tune.”

RAG (retrieval-augmented generation)

Fetching relevant documents first, then having the model answer using them.

Analogy
An open-book exam.
You’ll hear it
“The RAG pipeline retrieved the wrong policy.”

Source: Es et al. (2023)

Reasoning model

A model trained to work through a problem step by step before giving its answer.

Analogy
A student who shows their working before writing the answer.
You’ll hear it
“Use a reasoning model for the planning step.”

RLHF

Reinforcement learning from human feedback: improving a model using people’s rankings of its answers.

Analogy
A chef adjusting recipes from diners’ comment cards.
You’ll hear it
“RLHF is part of why chat models follow instructions.”

Source: Ouyang et al. (2022)

System prompt

Standing instructions given to the model before any user message, setting its role and rules.

Analogy
The employee handbook, read before the first customer walks in.
You’ll hear it
“Put the brand tone rules in the system prompt.”

Temperature

A setting that controls how varied a model’s word choices are. Higher means more varied.

Analogy
A creativity dial on the model’s word choices.
You’ll hear it
“Set temperature low for data extraction.”

Token

A chunk of text — often part of a word — that models read, write and bill by.

Analogy
Lego bricks of language.
You’ll hear it
“That long prompt eats tokens on every call.”

Tool use (function calling)

A model asking to run a defined function — search, calculator, database — and using the result.

Analogy
A worker reaching for the right tool in the toolbox.
You’ll hear it
“It called the search tool again with the exact same query.”

Vector database

A database built to store embeddings and quickly find the most similar ones.

Analogy
A librarian who finds books by topic, not by title.
You’ll hear it
“Store the help articles in a vector database.”

Vibe coding

Building software by prompting an AI and accepting its code with little or no reading of it.

Analogy
Driving by feel without looking at the map.
You’ll hear it
“Fine for a weekend prototype; not for payments.”

Source: Andrej Karpathy (2025)