Evaluation
also called Evals
How you know the system is actually working.
Systematic measurement of model or system quality on a defined set of tasks. It is the difference between "the demo looked good" and knowing whether a change helped.
Non-deterministic systems cannot be tested with exact-match assertions, so evals replace them. Build the eval set from real failing cases as they appear — that is where the value concentrates.
For agentic systems the definition widens: you are evaluating the model and the harness together, over multiple turns, with the final state of the environment as the outcome. A transcript that ends in a confident summary is not evidence; the files and tests are.
The demo that measured nothing
A prototype impresses three people in a review.
Write down twenty real inputs with the outcome you would accept, and run them. The number that comes back is usually lower than the room’s impression, and it is the only one that survives contact with users.
Evaluating only with examples you already tuned against. That measures memorisation, not quality.
- blogDemystifying evals for AI agents ↗
Gives working definitions for harness, task, trial, grader, transcript and outcome, and argues you must evaluate the model and harness together and judge final environment state, not just the last message.
- docsEvaluation best practices ↗
Evals as structured tests: task-specific test sets, continuous evaluation, logging, and human calibration of automated scoring.
- blogA statistical approach to model evaluations ↗
Why a benchmark-score difference may be noise: report uncertainty, account for clustered questions, use paired comparisons.
Benchmark
A shared public test used to compare models.
Golden Dataset · Eval set
A curated set of cases with known-good answers.
LLM as Judge · Model grading
Using a model to score another model’s output.
Regression Test
A test that catches a specific past failure coming back.
Eval Harness · Harness
The scaffolding that runs your test cases and scores the results.