Evaluations
An evaluation, or eval, is a repeatable measurement of what an AI system does on a defined set of inputs. Evals decide which model ships, which prompt change is an improvement and whether a safety mitigation works. Done badly, they produce confident numbers that do not survive a second run; done well, they come with sample sizes, intervals and a clear statement of what they cover.
Read the overview first. The benchmarks page explains what the major public benchmarks measure, why scores saturate and how contamination inflates them. The metrics page goes through accuracy, pass@k, calibration, judge-based scores and their failure modes. Statistical significance shows how many test cases you need and how to compare two models honestly. Evaluating reinforcement learning agents covers seeds, interquartile means and confidence intervals, which matter even more when training is noisy. The last page draws the line between red teaming and evals, and shows how they feed each other.
Our research notes, listed below, go deeper on judge error, clustered standard errors, contamination detection and more.
What are AI evals? A complete guide to LLM evals: types, components and how to build one
AI evals explained: capability, safety and product evals, the four components, sources of error, offline vs online evaluation, and how to build one.
Suggested reading order
- 1What are AI evals? A complete guide to LLM evals: types, components and how to build one
AI evals explained: capability, safety and product evals, the four components, sources of error, offline vs online evaluation, and how to build one.
- 2LLM benchmarks explained: what they measure, saturation, contamination and how to read leaderboards
What MMLU, GPQA, GSM8K, HumanEval, SWE-bench, IFEval and Chatbot Arena measure, why benchmarks saturate and leak, and how to read a leaderboard.
- 3LLM evaluation metrics: accuracy, F1, exact match, pass@k, ECE, win rate, BLEU and ROUGE
Accuracy, F1, exact match, pass@k, calibration error, win rate, BLEU and ROUGE: definitions, where each misleads, and which to use for which task.
- 4Statistical significance for LLM evaluation: paired tests, bootstrap, multiple comparisons and power
When is a gap between two eval scores real? Paired tests, bootstrap, repeated sampling, multiple-comparison corrections and power, with simulations.
- 5How to evaluate reinforcement learning agents: seeds, IQM and confidence intervals
Seeds, normalised scores, IQM, stratified bootstrap intervals and performance profiles, tested on 1,500 real Q-learning runs you can resample.
- 6AI red teaming vs evals: goals, methods, metrics and how they work together
Evals measure failure rates on fixed inputs; red teams hunt for failures adaptively. A side-by-side comparison, a simulation, and how to combine them.
Guides and explainers
LLM evaluation: the complete guide
Pillar guide.
LLM-as-a-judge: how it works, where it fails, how to validate it
Judge designs, documented biases, and a validation protocol with agreement metrics.
LLM-as-a-judge vs. human evaluation
Cost, agreement, latency and failure modes compared, and a hybrid design that uses each where it is strongest.
LLM evaluation plan template
A copy-ready plan: scope, threat model, categories, sample sizes, graders, gates and owners.
Research notes on this topic
How many test cases does an LLM eval need?
Wilson intervals, power analysis and a Monte Carlo study of paired vs. unpaired tests. Detecting a 2-point regression takes thousands of cases, not hundreds.
Your LLM judge is a diagnostic test: correcting measured failure rates
Sensitivity, specificity and the Rogan–Gladen correction. At a 0.5% true failure rate, a 97%-accurate judge reports 3.4%.
pass@k vs. pass^k: measuring agent reliability
Three task populations with identical pass@1 = 0.60 have pass^8 between 1.7% and 28.8%. How to estimate pass^k and how many trials you need.
Long-context evaluation beyond needle-in-a-haystack: tasks, position effects and effective context length
Multi-needle, aggregation and multi-hop tasks, position effects, precise definitions of effective context length, and how many trials it takes to measure it.
Rubric-based evaluation and rubric rewards: atomic criteria, weights and noisy graders
Designing atomic rubric criteria, weighting them, measuring per-criterion grader reliability, and using rubrics as RL rewards without criterion gaming.
Testing for sandbagging: elicitation, noise injection and the statistics of strategic underperformance
Testing whether a model hides capability on evals: fine-tuning elicitation, password-locked models, noise injection, consistency checks and statistical tests.
Predicting benchmark accuracy from small models, with intervals that cover
Sigmoid fits on log-compute, emergence as a metric artifact, continuous metrics and prediction intervals for scaling forecasts, tested by seeded simulation.
Evaluating tool-using agents: outcomes, trajectories, state diffs and pass^k
Outcome versus trajectory scoring, partial credit, state-diff checks for collateral damage, environment flakiness and why pass^k collapses with task length.
Know the number before you ship.
A 45-minute working session with an evaluation engineer. Bring one use case; leave with a threat model, a draft eval plan with sample sizes, and an honest read on where you stand.