TL;DR
- AI model evaluation answers one question, “is this model good enough for this job?”, by combining a metric, a set of tasks and a way of reading the scores. Change any of the three and the answer can change.
- There are four methods: automated metrics, benchmark suites, human review and LLM judges. Each catches failures the others miss, and the judge needs as much checking as the model.
- Everything here comes from Groundtruth, the benchmark we built for geological reasoning: eight models, 150 questions, one validated judge.
Key Takeaways
- A metric carries assumptions about the task. Word-overlap metrics assume one correct phrasing; open-ended answers break that assumption and get scored on the wrong basis.
- A public benchmark is a proxy, not your task. Across our three document sets, the lower half of the same eight-model table changes order from one set to the next.
- An LLM judge is part of the measurement. Ours must return a structured breakdown that a validator checks against the rubric; three failed attempts leave the question unscored rather than guessed.
- Every result has a noise floor. On 50 questions, the smallest real gap we can detect is 0.27 to 0.45 points, so a 0.3-point win is a finding on one set and noise on another.
AI model evaluation is the practice of measuring how well a model does a specific job, in a way that can be repeated and compared. Ask “is this AI model good?” and the honest answer is “for what, measured how?” A vendor quoting 83% on an academic benchmark has answered the second half of that question for a task you may not have.
We built Groundtruth to answer it for geology, and ran eight models through it: the same 150 questions, 50 per document set, graded by one fixed judge. The numbers in this guide come from those runs.
What is AI model evaluation?
AI model evaluation is the process of measuring a model’s performance on a defined task with a defined method (a metric, benchmark, human reviewer or LLM judge) so results can be compared across models or over time. What does model evaluation measure in AI? Never general intelligence: always one task, one reference, one scoring rule.
That narrowness is the useful part. A model can lead on one measured task and trail on another, and evaluating AI models well means stating which one you checked. On Groundtruth, Kimi K3 scored 8.81 out of 10 overall and GLM 4.7 scored 6.81; both numbers describe geological reasoning over a specific set of reports, and nothing else.
How to evaluate AI models: four evaluation methods
Most evaluation of AI models uses one or more of four methods. They differ in cost, speed and the kind of failure they can see.
| Method | Good at | Blind to | Cost per answer |
|---|---|---|---|
| Automated metrics | Tasks with one correct answer, at scale | Meaning, when an answer can be phrased several ways | Near zero |
| Benchmark suites | Comparing models on a shared, public task | Your own documents and failure modes | Low, once built |
| Human review | Nuance, relevance, subtle factual errors | Consistency across reviewers and scale | High |
| LLM judges | Rubric-based grading of open answers at scale | Its own biases, unless tested for them | Low to moderate |
Automated metrics
Automated metrics score output with a fixed formula. Accuracy works when there is one right answer. BLEU and ROUGE count the short word sequences an output shares with a reference answer, which suits translation and summarization. These AI model evaluation techniques are fast and cheap, and blind to meaning: a correct summary in different words scores badly, and a summary that copies the reference’s words with the meaning reversed can score well.
Our answers are open-ended geological reasoning, so a word-overlap metric would misgrade them. Our rubrics go further in the other direction: each lists accepted name variants so a spelling difference is not penalized, and a separate list of indicative terms that earn no marks on their own. Using the right vocabulary is not the same as making the right claim.
Benchmark suites
A benchmark suite is a fixed task set, usually with a public leaderboard, that tests a general capability such as coding or factual recall. It answers “how does this model compare with others on this task?”, not “will it work on my documents?”
We saw the gap directly. Groundtruth uses dynamic benchmarking to generate its questions from three separate document sets: Western Australian exploration reports, historical USGS ore studies and Canadian technical reports. The tiers held across all three sets, but inside the lower half the order changed from one set to the next. A single public benchmark would have hidden that. Building a benchmark from your own material is covered step by step in how to benchmark AI models.
Human review
A person reads the output and judges it. Human review catches what formulas miss: relevance, tone, subtle factual errors. It is slow, and two reviewers disagree unless the criteria are written precisely enough that both would mark the same answer the same way.
We use people where they add the most: writing the rubrics and the test answers used to check them, and resolving the cases a judge flags. The judge sets an adjudication flag when an answer may deserve credit it cannot verify, and every scoring run lists flagged questions for a person to review. The flag does not change the score; it marks where to look.
LLM judges
An LLM judge is one AI model scoring another model’s output against a rubric. It is the only method that grades open-ended answers consistently at scale, and it brings its own failure modes: it can favor the first answer it reads, favor its own model family, or produce a plausible score without applying the rubric at all.
Our judge, openai/gpt-5.5 at temperature 0, may reason at any length but must end with a JSON block: the gate decision, every component’s award and reason, the total, and an adjudication flag. A validator in the Groundtruth repository checks that block against the rubric: every component present, no invented ones, whole numbers within each maximum, a total that adds up, and zero everywhere if the gate failed. A block that breaks the contract is re-graded, up to three attempts, and then the question is recorded as unscored.
The full contract is in An LLM Judge That Can’t Invent a Score, and the wider method in our LLM-as-a-judge guide.
Choosing AI model evaluation metrics for your use case
Start with the cost of a wrong answer, not with the metric that is standard in your field. A spam filter and a model drafting a technical summary fail in different ways, and one accuracy number treats both failures as the same size.
Match the metric to the question
Groundtruth scores in two modes because they answer different questions. Pointwise scoring grades each answer 0 to 10 against its rubric and answers “how good is this?” Pairwise scoring shows the judge two answers and asks which is better, which answers “did this change help?” Only pointwise scores build our leaderboard: a pairwise verdict is relative, and swapping in a different baseline can change every verdict.
Pairwise has a known trap, position bias. We ask the judge twice with the answers swapped and average the two verdicts, so a judge that always prefers whatever it reads first produces a tie instead of a false winner. The mechanics are in Position Bias in Pairwise LLM Evaluation.
Gate the score
A rubric without a gate pays partial credit to fluent answers that miss the point. Each of our rubrics has one gate component worth 2 to 4 points: miss the claim the question exists to test and the whole answer scores zero. Only after the gate passes do the other components add points.
Know the resolution before reading a gap
On a 50-question set, the smallest mean gap we can tell apart from noise was 0.27 points on the Western Australian set, 0.33 on USGS and 0.45 on the Canadian technical set. A 0.3-point improvement is real on the first and noise on the third, from the same benchmark and the same judge.
Put cost next to quality
Generating answers for all 150 questions cost $7.47 with the cheapest model and $54.97 with the most expensive, and the most expensive did not win. Once the top models sit inside each other’s noise floor, as our top three do, price is often the only difference left to act on. Current scores and costs are on the Groundtruth leaderboard, and what they mean for model choice is in Which AI Model Is Best for Your Geology?
AI model evaluation tools
Off-the-shelf AI model evaluation tools fall into four categories. None removes the choices above.
| Category | What it gives you | Examples |
|---|---|---|
| Eval libraries | Metrics and assertions that run like unit tests in CI | DeepEval, OpenAI Evals, Ragas for retrieval |
| Research frameworks | Sandboxed, reproducible runs with a task, solver and scorer | Inspect AI |
| Leaderboards and tracing platforms | Public model comparisons; traces and datasets from production traffic | Artificial Analysis, LangSmith, Arize Phoenix |
| A custom harness | Your own tasks, rubric, judge and isolation | Groundtruth, in our case |
Libraries suit teams that want evaluation to fail a build. Tracing platforms suit teams whose main risk is production behavior they cannot see. A custom harness is worth building when the reasoning you need to measure has no general metric, which was our case for geology.
Common mistakes in evaluating AI models
Trusting a judge that has never been tested. Before a rubric grades any model, we run fabricated answers with known expected scores through it. On two benchmarks, 14 and 16 of those test answers scored above the author’s prediction, and the cause was the same both times: two components rewarding the same idea. On one project the rubric’s author found 2 defects by reading; the test answers found 15 more. The method is in Testing the Rubric Before You Test the Model.
Checking structure but not judgment. A valid JSON block can still hold a wrong decision. During calibration, one fabricated answer reached the right conclusion by an argument the question was built to reject, and the judge gave it 10 out of 10 against an expected 0.
Letting the model see the answer key. When the model under test is an agent with file tools, the rubric can sit inside its search space. With only the obvious hiding spot closed, one of our 50-question runs reached the rubric folder on 13 questions. We now move grading files out of the repository for the run and audit every transcript before its score counts. The full account is in Your AI Agent Can Read Its Own Answer Key.
Ignoring the judge’s own conflicts. Our judge comes from a model family that is also on the leaderboard, where GPT 5.6 Sol placed third. We disclose that in the methodology rather than leaving readers to find it.
Reading rank order as a result. Below the noise floor, a rank swap between two models is not a finding. Report tiers, not places, when the gaps are that small.
FAQs
What does AI model evaluation measure?
AI model evaluation measures one capability on one task: accuracy against a reference answer, a score against a rubric, or a preference over an alternative, with cost and consistency alongside when they matter. It never measures general intelligence. On Groundtruth, it measures geological reasoning over specific reports, scored 0 to 10 per answer.
How do you evaluate an AI model?
Define the task and what a wrong answer costs. Pick a method that fits it: automated metrics for one-answer tasks, benchmark suites for general comparison, human review for nuance, or an LLM judge with a rubric for open answers at scale. Test that method on answers with known scores before trusting it, and check the result’s noise floor.
What metrics are used to evaluate AI models?
Accuracy for tasks with one correct answer; BLEU and ROUGE for translation and summarization; rubric scores, such as 0 to 10 per answer with a gate, for open-ended reasoning; and pairwise preference for comparing two versions. Report cost alongside quality, and state the smallest score gap your sample size can reliably detect.
What tools exist for AI model evaluation?
Eval libraries such as DeepEval and OpenAI Evals run checks in CI. Research frameworks such as Inspect AI give reproducible sandboxed runs. Leaderboards such as Artificial Analysis compare models publicly, and tracing platforms such as LangSmith build test sets from production traffic. A custom harness covers domain tasks no general tool measures.