You will pull the same fields out of four contracts, and then answer a harder question: how often was it right? Not "does that look correct". A number.
Eight parts. The first three look like ordinary work, and they are. The rest is what turns that work into something you can trust.
Twenty to thirty pages each, written the way a real supplier contract is written. Every sector signs these: law, accounting, financial services, personal injury.
A schema is the formal definition of the data you want out. Three things for every field.
"Contract value" sounds obvious until you open the document and find three numbers that all have a claim to it. Write down which one you mean.
Point Claude Code at your schema and at the Northgate agreement. One row per field, following the match rule column.
Look it over. Does anything seem off to you?
Hold that. Do not fix it yet. We will come back to it with a number instead of an opinion.
Turn the working prompt into a file so you can run it again without retyping. Ask Claude Code to write it, and to ask you clarifying questions first.
Nothing clever. The reason we are saving it now is that in ten minutes you are going to want to run this same extraction four times, and then run it again after you change it.
Run the skill across all four. One record per agreement, plus the complete charge schedule for each. Not a sample. All of it.
Try eyeballing that. Tell me your error rate.
By hand, checking each value against the document. That is the answer key.
Ask Claude Code to compare its output to the answer key, field by field, and report where it matches and where it misses.
Pick one of the failures and read both values out to me.
A correct answer can look different from the answer key. Strip the formatting off both sides before you compare.
Normalise the formatting on both sides before you score, or you will grade the punctuation instead of the data.
Everything still failing after normalisation traces back to one field whose match rule was vague, or blank. Open the schema you filled in an hour ago.
You did not write a data definition this morning. You wrote the grader, and then you wrote a tool that gets graded by it.
The specification and the answer key are the same document.
A wall of pass and fail text is technically an eval and practically useless. You want to look at this and know within three seconds which field is your problem.
Take the top failing field. Work out which of three things it is, change the rule, and re-run the whole thing.
Correcting the answer fixes one cell today and nothing tomorrow. You change the rule and re-run. If the number goes up, the rule was the problem. If it goes down, you have learned something about your own thinking.
A number you trust too much is worse than no number. Five things this class did not teach, and you should know all of them.
What it gives you is direction. Change a rule, and within a minute you know whether it helped or hurt. That is enough to stop guessing, and stopping guessing is most of the value.
An answer key grades extraction, where one value is right. A rubric grades judgment, where several outputs could be good. Between them you can measure both halves of the work. Neither one measures the other.
tutto.one/praxis · daniel@tutto.one