Measure the output

Now run the pipeline and collect the numbers. The discipline here is simple and mostly ignored: measure with the same ruler every time, and write the results down.

What to record

For each run, capture:

The setup — dataset version, model, quant, runtime, settings, prompt version. Without these, the score is meaningless.
The raw scores — pass/fail per example, or the scale you defined.
The failures — not just the count, but which examples failed and how. The failures are the useful part.

A single aggregate number hides the story. "17/20" tells you little; "failed the three examples about dates" tells you where to look.

Metrics that mean something

Pass rate — the headline number, but only against your fixed criteria.
Error type — did it fail by omission, by wrong fact, by wrong format? Group the failures.
Consistency — run the same example a few times and see if the answer is stable, especially if you use a nonzero temperature.

Do not over-metric a small personal bench. A pass rate plus a short list of failure types is plenty, and it is honest.

Write it down

A result that lives only in your head is gone in a week. Put every run in the same place: a file, a table, a changelog. Over time that table is your expertise, because it shows what actually moved the numbers and what did not.

The connection forward

Measuring is only half the loop. The next lesson is the other half: change one thing, run again, and compare. The point of measuring now is to make that comparison possible, and to make it honest.

A number you recorded is a fact. A number you remember is a guess. The method is built on facts.