What a Translation Score measures, and what it does not

One number, built from findings that already exist

The score triggers no model call of its own. It is an aggregation of findings the Judge has already produced for each field: category, severity, reasoning, kept for 90 days. Calculating the score costs nothing.

For any set of checked fields, a job, a language, a whole shop or a period, the score is 100 × (1 − Σ gᵢ / n). Here n is the number of checked fields and gᵢ is the weight of the most severe finding in field i. A field without findings has gᵢ = 0. Several findings in one field count once, with the highest weight, so a single broken field cannot dominate the result.

The weights

Four of the five heaviest finding types come from deterministic checks that cost nothing. A quick check without any model call is therefore already a solid base for the score. The Judge refines it.

Weights are versioned. A score always states which version produced it. Changing a weight creates a new version, old scores are not recalculated.

What the score cannot tell you

It cannot tell you whether a translation sells. Tone and style carry a weight of 0,15 because they are a matter of taste, not an error. It cannot replace a native reader for legal texts. And it is only as good as the sample it was calculated on.

Sample size is part of the number

A score from 12 fields is not the same as one from 12.000. Below 30 fields Glotari shows the score in grey, marked preliminary, with no band. Between 30 and 199 fields the band appears together with the sample size. From 200 fields the addition disappears. No confidence interval, the sample size is enough.