Scores Guide

You already have standards written down, and text that has to clear them:

• Marketing copy that has to match your brand voice before it ships.
• A legal or compliance line that has to hold up against the regulation it cites.
• A support answer that has to match the policy in your help docs.
• Several drafts of the same thing, when you need to pick the strongest one.

A scores run grades that text against your rubric or compliance store and tells you, criterion by criterion, where it passes and where it falls short, quoting the exact phrase behind every mark.

Scoring against a rubric needs Pro; scoring against a compliance store needs Pro+.

This guide covers how to score a single piece, how to rank several at once, and how a saved standard lets you run the same check from the dashboard, the API, or an agent, without leaving where you work.

Running a scores run

You give it two things: the text to judge and the standard to judge it against.

• Artifact. The text you want scored. Paste the finished copy, answer, or clause.
• Rubric. The named criteria the judge scores against. Required unless you are scoring against a gold-standard answer instead.
• Compliance store. Optional, Pro+. Pairs with a compliance rubric so each criterion is scored against the clause retrieved from your document, not the criterion text alone.
• Gold standard. An alternative to a rubric: paste the ideal answer and the judge scores semantic fidelity to it.

You get back a score and the judge’s reasoning for every criterion. It grades the exact words you gave it, so the verdict is about the text you are shipping, not a regenerated version of it.

The score sits beside the exact phrase that earned it, criterion by criterion

See the Rubric Guide for writing criteria that score reliably, and the Compliance Guide for structuring a compliance store.

Score one piece, or rank several

Hand a scores run a single artifact and you get one scorecard. Hand it several and you get each one scored against the same standard, ranked. This is how you pick the strongest option without arguing about it: write three versions, score them against the rubric that encodes what good looks like, and read the result.

Here are three candidate taglines, scored against a brand-voice rubric that rewards a concrete reader benefit and penalises unbacked hype and filler.

One standard, three candidates, scored the same way. The ranking is the argument.

The judge quoted the exact phrase behind each loss. The bottom tagline scored 0 on No unbacked hype for “most accurate”, a superlative the rubric forbids. The middle one scored 0 on Concrete and specific for the filler words “next-generation” and “seamless”. The winner led with a reader action and lost points only where the judge read it as closer to a sentence than a tagline. Once the rubric is written down, the ranking is evidence, not opinion.

Reading a scores result

A scores result shows the overall score, then each criterion with its own score and the judge’s reasoning. Every number traces to a line of reasoning, so a low score always comes with the phrase that caused it. For a compliance run, the result also shows which clause the judge retrieved for each criterion, so the verdict is traceable to a specific passage in your document.

See the Results Guide for how scores are calculated and how to read the diagnostic findings that come with a run.

Turn a standard into a preset, then check anything

Save the rubric, or the compliance rubric and its store, as a preset, and the standard becomes a named thing anyone can run: “marketing-standard”, “legal-review”, “compliance-v2”. With that preset reachable over MCP, “check this against our marketing standard” becomes one sentence in a chat window, your IDE, or a CI step.

One message, scored against the standard the team already saved, with the exact phrase at fault and the criterion it failed.

See the Presets Guide for saving a standard as a preset, invoking it across all three surfaces, and running it on a schedule to watch a standard over time.

What’s next