Eval Scores Chart

About this block

Eval Scores Chart treats AI quality like a metric instead of an adjective. Two lines track releases, the share of drafts editors accept unchanged rising, the share flagged by fact checking falling, and the footnote owns the one release that regressed. Publishing the bad release is the move, it makes the good releases believable.

The chart is built on the shadcn chart primitives, so the data array at the top of the file is the only thing to replace when wiring your real eval history. Config labels and grayscale chart tokens come from chartConfig.

When to use

Reach for this block on AI pages aimed at buyers who have heard every accuracy claim already. It assumes you actually run evals per release, do not fake this one.

A natural flow around it on a Marketing Pro page:

Before

• Drafts demo showing the workflow the evals measure

After

• AI principles covering training data and human review policy
• Trust where the rubric is documented

What you can display

One strong use is acceptance rate and factual flag rate per model release. Other evals:

• Latency per release. P50 and P95 generation time.
• Coverage. Share of briefs the assistant attempts versus declines.
• Benchmark suites. Public benchmark scores across versions.

Tip: include the regression release. A monotonically improving chart reads as marketing, a dented one reads as measurement.