The results screen. Per criterion first, then the cases — the single number at the top is the least useful thing on the page.
The problem
You changed the prompt. Did it help? That question usually gets answered by trying three examples and forming an impression — which is not an answer, because the next person to ask will try three different examples and form a different one.
The thing missing is not a model. It is an instrument: a fixed set of real inputs, a written definition of good, and a number computed the same way every time so that two runs can be compared at all.
What it does
Cases
Real inputs, not invented ones. Twelve support messages for an online store, of the shape a helpdesk actually receives.
Rubric
What good means, in language precise enough that two people reading it reach the same verdict. Four criteria, three of them blocking.
Run
Score every case against every criterion, by model or by hand, and keep the run.
Diff
Which cases moved since the last run, and in which direction. This is the screen the whole thing exists for.
What it showed
Overall
76%
Across twelve graded cases. A respectable pass, and on its own it tells you nothing.
Release blockers
5
Cases that failed a blocking criterion. Nearly half the set is unshippable whatever the score says.
The two numbers are the finding. Seventy-six per cent is the sort of figure that survives a steering meeting; five blockers out of twelve is not. The average quality is fine and the floor is not, and only one of those facts reaches you if the tool reports a single score.
Broken out by criterion, the run stops being a grade and becomes a work queue: states policy correctly is at 75% against 83% for the other two blockers, and all three of its failures are the same class of error. Tone is the weakest score on the page at 63% and the least urgent thing on it.
What it cannot do
It grades what you bring. A test set of fifty comfortable cases produces a comfortable number, and nothing in the tool can tell you that the tail you did not sample is where the failures live.
If a model does the grading, the grader needs auditing too — which is why there is a tab for exactly that: score a sample by hand, compare, and treat every disagreement as a hole in the criterion rather than a fault in the model.
The long version of the argument behind it is The rubric that replaced the spec.
The tool
Open Rubric
The working thing, with the example above already loaded — every tab live, the walkthrough on the first screen, and nothing you type leaving your browser.
See the prototype →