Evals and measurement is the practice of finding out whether an AI system is right, using evidence that does not come from the system itself. It sounds obvious. In practice, most AI dashboards report proxies: approval rates, thumbs-up counts, a judge model's pass rate, or a single headline score on the items the prompt was tuned against. Those numbers move, but they don't tell you how often the system is correct, on which kinds of work, or how sure you can be.
The essays here share one method with a few fixed parts:
- Start from one business outcome. Write down both the success you want and the error that must never happen.
- Build the answer key before you see model output. Use real cases, labelled blind by two people against a versioned guideline, and check their agreement with kappa.
- Evaluate every step of a multi-step flow. Success compounds: five steps at 95% each complete about 77 runs in 100.
- Report every rate as a range and gate on its lower bound. 49 of 50 only supports about 89.5%.
- Give severe errors their own gate. Overall gains never offset them.
- Keep the measurement honest. Hold out a set nobody tunes on, split results by slice and by action, measure consistency across repeated runs, and check any model judge against people before trusting it.
These essays are written from the operator's side. They cover the arithmetic, the traps, and the judgement calls about when the evidence is strong enough to let a system act alone and when a slice should stay with people. One idea runs through all of it: trust comes from measurement that is independent of the system being measured.
Start here
- Your answer key is the product: why the labelled set matters more than the model.
- A rate is a belief with a width: reading every pass rate as a range.
- Counting honestly: ranges, zero failures, paired changes: sizing tests and comparing versions.
- Pick the metric by what a mistake does next: choosing precision or recall for each action.
- Counts out of a thousand: what a flag really means: base rates in whole numbers.
- A 99 percent approval rate is not an accuracy number: proxies versus correctness.
- A lucky week is not a better model: noise, luck and regression to the mean.
- Why agents cannot grade their own work: why the grader has to be independent.
- Agents need an independent exam: testing agents against evidence they did not produce.
- Model consensus is not engineering evidence: why agreement between models is not proof.
Questions this topic answers
- Why do our evals pass while production still surprises us?
- How do we turn a business KPI into a gold set and per-step checks?
- How many test items do we need before an AI step can act on its own?
- What does a pass rate mean when the sample is small?
- When can we trust an LLM judge?
- Which metric fits an action: precision, recall or something else?
If you are building measurement into your own AI work, connect with me on LinkedIn to compare notes and discuss.