Build a team with the right skills
One person should not design, build and judge the eval alone. Someone who knows the domain decides what good looks like; someone who knows methods decides how to sample and score.
Anthropic published ten recommendations for designing, running and reporting evals. Here they are as a checklist you can tick against your own eval, each one tested on a real Claude Code eval of ours.
Read the original: “Building strong safety evaluations”, Anthropic, 13 November 2026Start here
If you only do three things, the post says to:
“write down the construct you want to measure before collecting data, validate your grader against human judgment before trusting any number it produces, and report uncertainty so readers can tell signal from noise.”anthropic.com/research/building-strong-safety-evaluations →
We graded our own Fable 5.2 vs Opus 5.5 Claude Code experiment against all ten rules. It is a coding and cost eval, not a safety eval, which is exactly why it is a useful test: the rules hold up outside safety work too. Our eval scored:
What held up once we added ranges: Fable cost 2.0x to 2.68x as much as Opus for the same result. What did not: the speed difference is within noise, and the 6/6 quality tie only says both pass rates sit somewhere between 54% and 100%. Details are under rule 8.
Tick each question that is true for your eval. A rule counts toward your score when all of its boxes are ticked. Your ticks are saved in this browser only.
Your score 0 / 10
One person should not design, build and judge the eval alone. Someone who knows the domain decides what good looks like; someone who knows methods decides how to sample and score.
Assume the model will try things you did not plan for. Give it a real sandbox, cut its network, and watch runs live so you can stop one.
A plain score comparison tells you that two models differ, not why. If you want to claim a cause, change one thing at a time and hold the rest fixed.
Your metric is a stand-in for the thing you care about, and stand-ins drift. A model can top the metric while missing the point.
A few easy, single-shot tasks rarely predict real use. Vary the things that matter one at a time, and test in the setup you actually ship.
Close every shortcut to full marks, make sure a known-good answer passes your grader, and have the grader fail loudly when it breaks instead of quietly scoring.
Decide the metric, the comparison and what counts as a real difference before you see results. Plan repeat runs. Label everything else as exploratory.
Every number needs a range. If the range is wide, a gap between two models may be noise, and the claim has to say so.
Write down enough that a stranger could rerun it: model versions, settings, prompts, grader version and dates. Never compare scores from different graders.
Report what went wrong alongside what went right. If you find a serious flaw in someone's shipped system, tell them first and give them time to fix it.
Rule 8 in one box. Enter how many runs passed out of how many you ran, and get the range the true pass rate probably falls in. Try 6 of 6: a perfect score from six runs still leaves a lot of room.
Observed 100.0%. The true pass rate is plausibly between 54.1% and 100.0%.
95% Clopper-Pearson interval. Computed in your browser, nothing is sent anywhere.
An eval is a repeatable test for a model or an agent: a fixed set of tasks, a way to score each answer, and a report of the results. It is how you tell whether a model change, a prompt change or a new model actually made things better, instead of going on a feeling from a few chats.
Anthropic treats 100 items as the floor for a complex, multi-turn eval and 1,000 for a simpler single-turn one. You can learn from fewer, but with a handful of cases the range around any score is so wide that most differences are noise. The calculator on this page shows how wide: 6 passes out of 6 runs still only tells you the true rate is somewhere between 54% and 100%.
It is the range of true values that are consistent with what you observed. If you report "95% pass rate" from 20 runs, the real rate could easily be several points lower or higher. A 95% interval gives that range. When two models' ranges overlap heavily, you cannot say one is better.
LLM-as-judge means using a model to grade other models' answers against a rubric. Check it before you trust it: grade a sample by hand, compare the judge's labels with yours using an agreement measure corrected for chance, use a different model from the one being tested, hide which model wrote each answer, keep each criterion yes/no, and look for a bias toward longer answers.
No. Anthropic wrote the guide for safety evaluations, but every rule applies to any eval: coding, cost, retrieval quality or customer support. Construct validity, grader checks and honest ranges matter just as much when the question is "which model is cheaper" as when it is "which model is safer".
Both are open-source tools named in Anthropic's post. Inspect, from the UK AI Security Institute, is a framework for writing and running evals with logging built in. Petri is a tool for automated auditing that runs multi-turn conversations with a model to probe its behaviour. Either one saves you from hand-building the harness and records much of what rule 9 asks for.
We graded our own Fable 5.2 vs Opus 5.5 Claude Code eval against each rule on 14 November 2026, using its results files, scripts and raw run logs. The tally was 3 pass, 5 partial and 2 fail. We graded our own work, so treat the grades as a worked example, not an audit.
Building evals or AI systems at your company and want expert help? Eight Labs consults for enterprises, SMEs and startups. Get in touch.