BlogLLM Evaluation
// guide
LLM evaluation: measure a prompt or agent before you ship
LLM evaluation is the practice of scoring a prompt's or an agent's outputs against a fixed dataset and a defined standard of good, so a change is measured before it ships and watched after it does. It replaces "it looked fine in testing" with a pass rate, a trust score and a list of exactly what regressed.
Unit tests cannot do this job. The same input produces different outputs on successive calls, the provider can change the model behind a fixed API version, and the thing being judged is quality rather than equality. So evaluation needs its own machinery: a dataset of real inputs, scorers that are deterministic where they can be and model-graded where they must be, thresholds calibrated to the variance you actually observe, and a place in the pipeline where a regression informs the person who decides what goes live. The articles below cover each part, from what an eval is to how an LLM judge goes wrong and how to catch a hallucination before a user does.
The numbers behind the discipline: Anthropic's August 2025 postmortem documented routing errors affecting up to 16% of Claude Sonnet 4 requests with no API modification, which is the "something changes you" case no unit test sees. OpenAI's April 2025 o3 system card documented o3 hallucinating on 33% of PersonQA prompts versus 16% for its predecessor o1, so a more capable model is not a safer one by default. And on the judges themselves, research found that padding a response with a redundant summary paragraph can shift an LLM judge's score by 0.5–1.5 points on a 5-point scale.
Articles on LLM evaluation (9)
LLM Hallucination Detection: How to Build a Detector
Implement LLM hallucination detection at eval, trace, and runtime. Score for groundedness, flag live failures into tests, and roll back in one command.
· 8 min read
The Future of Agentic Hallucination in Production
Hallucination rates don't improve linearly with model size. Here's where the problem is heading as agents compound errors — and how to gate against it.
· 6 min read
Eval Metrics: What to Measure, What to Gate, What to Skip
Which eval metrics to run in CI, how to calibrate thresholds, and the agentic failure modes most guides skip.
· 6 min read
How to Test Prompts: Assertions, Thresholds, and CI Gates
Write prompt tests that catch real regressions: test case structure, scorer selection, golden datasets, and five failure modes every team hits.
· 7 min read
LLM as Judge: How It Works and When to Use It
LLM as judge uses one model to score another's outputs. Learn how it works, the three judging modes, common biases, and how to wire it into your CI/CD pipeline.
· 7 min read
What Are LLM Evals? Dataset, Task, Scorer, CI Gate
LLM evals are automated tests that measure whether your AI prompt or agent does what you intend—dataset, task, scorer—and how to gate CI/CD on the result.
· 8 min read
AI regression testing in CI/CD
Four changes reliably break AI features: prompt edits, model swaps, provider drift, and schema changes. Here's the CI test suite that catches all four.
· 6 min read
LLM Evaluation Platform: What It Is and What It Does
An LLM evaluation platform runs automated quality checks on AI outputs—offline in CI and online in production—to catch regressions before they reach users.
· 7 min read
Testing Prompts Before Shipping: Dataset, Threshold, CI Gate
Testing prompts before shipping means running them against a fixed dataset and failing CI when the pass rate drops below your threshold. Here's how.
· 6 min read
FAQ
- What is LLM evaluation?
- LLM evaluation is the practice of scoring a prompt's or an agent's outputs against a fixed dataset and a defined standard of good, so a change is measured before it ships. An eval has three parts: a dataset of inputs, a task that produces outputs, and scorers that turn each output into a number your pipeline can act on.
- How is LLM evaluation different from software testing?
- A unit test asserts that an input produces exactly one correct output. An eval measures whether outputs are good enough across a distribution, because the same prompt gives different answers on successive calls and quality is a judgement rather than an equality. Evals also catch the regressions you did not cause, such as a provider changing the model behind a fixed API version.
- Can an LLM be used to grade another LLM's output?
- Yes, and for open-ended quality it is often the only practical scorer. An LLM judge scores or ranks outputs against a rubric. It carries known biases, including position, verbosity and self-preference, so treat its scores as a signal, run pairwise comparisons both ways, and keep a human review cadence as the calibration layer.
- Do evaluations block deploys on Functional AI?
- No. Deploying runs the Contract Test, an instant structural check of the definition, and nothing else. Experiment results, including a trust score and findings with a fix for the prompt, inform a human who decides what goes live. fn eval exits non-zero on failure, so a team that wants a hard bar can run it in its own CI pipeline.
Evaluate a prompt or agent on your own data
Functional AI is an LLM evaluation platform built into the platform that runs your Agentic Function: datasets, Score Sets, experiments across models, a trust score, findings with prompt fixes, and behaviour tests generated from an agent's own definition.
See evaluation on Functional AI →