A quality score means nothing unless you know how it was produced. Here is ours. If you use a different framework, we will work in yours — say so and we will tell you what changes.
The error types
We classify every error by type. The types are those of MQM, the published multidimensional quality framework that most buyers in this industry already use, so our numbers can be compared to yours.
| Type | What it covers | |—-|—-| | Accuracy | The translation does not say what the source says — mistranslation, omission, addition, untranslated text | | Terminology | A required term is not used, or is used inconsistently | | Linguistic conventions | The target language itself is wrong — grammar, spelling, punctuation, word order | | Style | Grammatically fine, but wrong in tone or register, or simply awkward | | Locale conventions | Dates, numbers, currency, addresses, telephone formats, units | | Audience appropriateness | Technically correct but wrong for the intended reader | | Design and markup | Tags, formatting, layout, text overflow, right-to-left mirroring |
The severity levels
Every error also gets a severity, and severity carries a weight. The weights rise steeply, on purpose.
| Severity | What it means | Points | |—-|—-|—-| | Neutral | Flagged for attention, but acceptable. A note, not an error | 0 | | Minor | Noticeable, but does not impede understanding or use | 1 | | Major | Seriously affects whether the content can be understood, trusted or used | 5 | | Critical | Makes the content unfit for purpose, or creates a risk of physical, financial or reputational harm | 25 |
The jump from 5 to 25 is not arbitrary. Twenty-five punctuation slips are annoying. One wrong dosage figure is a different kind of event, and a scoring system that treats them as comparable is counting typos rather than measuring quality.
And one rule that sits outside the arithmetic: any single critical error fails the file, whatever the score. A good average does not rescue a dangerous document.
The arithmetic
Total points = (minor × 1) + (major × 5) + (critical × 25)
Penalty points per 1,000 words = (total points ÷ words in sample) × 1,000
Score out of 100 = (1 − (total points ÷ words in sample)) × 100
We lead with penalty points per 1,000 words and report the score out of 100 second, for a practical reason: professional work has few errors relative to its word count, so the score out of 100 compresses nearly everything into the range 97 to 100, where differences are invisible. Points per thousand words spreads the same information out where you can read it. 12 is clearly better than 46.
The threshold, agreed before we start
The pass mark is not the same for every kind of content, and we would rather agree it with you in advance than discuss it once the number exists. A starting set:
| Content type | Pass at or below, per 1,000 words | |—-|—-| | Regulated — medical, legal, safety, chemicals | 10 | | Customer-facing — product, support, web | 20 | | Marketing and campaign | 20, with style errors weighted higher | | Internal or low-stakes | 40 |
Plus the critical-error rule, which overrides all of them.
If the threshold is agreed beforehand, the score is a finding. If it is agreed afterwards, it is a negotiation. That is most of the difference between an assessment and an opinion.
What we do not count
If a reviewer cannot name a type and a severity, it is not an error — it is a preference. Preferences are logged in a comments column and never counted in the score. Translators find this genuinely difficult, and it is the first thing we train for.
Sampling
We score a defined sample, and we say in the report how the sample was chosen — random, or targeted at the highest-risk sections. Both are legitimate. Not saying which is not. Below about 500 words a single bad sentence swings the whole number, so we use 1,000-word samples where we can and flag it in the report when we cannot.
Your framework, not ours
Where you use DQF-MQM, or J2450 in automotive, we work in your framework and report in your categories. It costs us nothing and it means you do not have to translate our numbers into yours.